Academic Databases
Explore Our Semantic Databases
STEM & Tech
Math, Physics, CS & Engineering...
Economics
Macro, Micro, Econometrics...
Business
Management, Marketing, Finance...
Health & Life Sciences
Biology, Medicine, Neuroscience...
Humanities & Social Sciences
Psychology, Law, Philosophy, History...
General Studies
General & interdisciplinary studies...
Explore Topics
Features

Fine-Tuning vs RAG vs AI Agents: A Comparative Analysis

Added: 2026-09-10

Deep Dive

Fundamental understanding of Large Language Models (LLMs), including how they are pre-trained and their inherent limitations like hallucinations and knowledge cutoffs.

6. An Introduction to RAG - Part of the Free Ollama Course

6. An Introduction to RAG - Part of the Free Ollama Course

39K

AI models have inherent weaknesses including hallucinations (generating incorrect information) and knowledge cutoffs. Models cannot be easily told not to hallucinate because they don't always recognize when they're doing so. Additionally, models cannot guarantee consistent good answers to the same question across different instances. Models typically have knowledge cutoffs because training takes months, meaning their last known information is from several months ago unless fine-tuned. Fine-tuning doesn't add much recent factual knowledge and can slow down the model's learning of new facts.

Click to watch →

LLMs are like your weird, over-confident intern | Simon Willison (Datasette)

LLMs are like your weird, over-confident intern | Simon Willison (Datasette)

3.2K

LLMs have significant limitations: (1) They cannot count reliably, (2) They hallucinate confidently about internet-famous people by generating statistically likely text from training data, (3) They cannot access the web without browsing tools, and (4) They produce fake URLs that look real but aren't clickable. Understanding these limitations is crucial for responsible use. The key is recognizing that LLMs are not knowledgeable but are generating statistically likely text from their training data.

Click to watch →

DevCoach 238: ML | Mengenal RAG: Dari Konsep ke Implementasi Pertamamu

DevCoach 238: ML | Mengenal RAG: Dari Konsep ke Implementasi Pertamamu

464

Large Language Models have three fundamental limitations: (1) Context limitation - cannot access private documents users possess, (2) Knowledge cutoff - only contain information up to training date (e.g., October 2023), and (3) Hallucination - generate convincing but factually incorrect answers. Fine tuning retrain LLMs with specific data but requires large labeled datasets, significant GPU resources, and lacks flexibility for document changes. RAG (Retrieval Augmented Generation) addresses these limitations by connecting LLMs with external documents through three stages: Retrieval (finding relevant documents), Augmentation (adding retrieved information to context), and Generation (producing answers based on context).

Click to watch →

LlamaIndex Webinar: Learn about Fine-tuning + RAG (w/ Victoria Lin, author of RA-DIT)

LlamaIndex Webinar: Learn about Fine-tuning + RAG (w/ Victoria Lin, author of RA-DIT)

4.8K

This comprehensive section covers the theoretical foundations and technical implementation of RA-DIT. It begins by explaining why pure parametric language models face limitations: knowledge cutoff after pre-training, hallucination of rare facts, and inability to show provenance. The evolution of retrieval-augmented approaches is traced through key milestones: Retrieval-Augmented Generation (RAG) using dense passage retrieval for end-to-end training; RRoT architecture from DeepMind coupling retriever memory with language models through cross-attention during pre-training; and Replug, a simple methodology augmenting black-box models with retrieved content without custom architectures. The RA-DIT framework is introduced as addressing how to strengthen coupling between language models and retrievers through fine-tuning, operating in two distinct steps: language model fine-tuning maximizes likelihood of producing correct answers using all retrieved passages, promoting robustness against noisy retrieval content; retriever fine-tuning uses language model prediction probabilities as feedback signals, minimizing KL divergence between language model scores and retriever output probabilities. The experimental methodology section details the technical implementation: using Llama series models (primarily 65B parameters) and state-of-the-art Dragon retriever; fine-tuning using standard supervised instruction learning with label loss computed only over task outputs; parallel retrieval augmentation creating multiple augmented prompts by generating multiple retrieved passages (3 during fine-tuning, 10 during inference); and careful task selection promoting knowledge utilization and contextual awareness across reading comprehension, summarization, open-domain QA, dialogue assistants, and chain-of-reasoning datasets.

Click to watch →

What's AI good for? (Probably not that, but not nothing)

What's AI good for? (Probably not that, but not nothing)

303

This section covers critical limitations of LLMs. They have no inherent knowledge of current date, time, or external information—everything must be in context. LLMs can use tools through prompting, where the model generates tool requests that external systems execute. Thinking tokens allow models to emit intermediate content into their own context, improving performance but not representing genuine 'thinking.' Hallucinations occur because LLMs fill in information they don't know, similar to human confabulation. The training data doesn't fit entirely in the model (like JPEG compression), so the model must fill gaps based on logical patterns. This is not unique to AI—humans do the same thing.

Click to watch →

Deep Dive

Basic concepts of Information Retrieval (IR) and semantic search, particularly how text embeddings and vector databases function.

Vector Database Explained | What is Vector Database?

Vector Database Explained | What is Vector Database?

238.5K

Vector databases enable semantic search by converting text into numerical representations called embeddings. Unlike keyword matching, semantic search understands query intent and context. Word embedding assigns numerical values to words based on properties like category, location, or revenue. Similar words have similar vectors, enabling similarity calculations. Advanced techniques like word-to-vector arithmetic allow mathematical operations on word representations. Transformer-based methods, used in modern AI models like ChatGPT, generate sophisticated embeddings through complex statistical techniques.

Click to watch →

Ce plombier a un meilleur système IA que la plupart des devs

Ce plombier a un meilleur système IA que la plupart des devs

184

Vector databases store text as numerical vectors (embeddings) representing semantic meaning. Similar concepts have similar vectors (dog and puppy are close), while unrelated concepts are far apart. This enables powerful semantic search where 'old dog' might find 'puppy' even if exact words don't match. The system converts text to vectors, stores them in databases, and retrieves similar vectors when searching, making information retrieval much more powerful than keyword search.

Click to watch →

CODE Presents: Intro to AI with MS Semantic Kernel

CODE Presents: Intro to AI with MS Semantic Kernel

1.8K

Vector databases store text as numerical vectors generated by embedding models, enabling semantic similarity searches rather than keyword matching. Text-embedding-ada-002 converts text to dense vectors where semantically similar texts produce similar representations. Collections organize stored information with unique identifiers. Searches return results based on mathematical closeness of vectors, solving the problem of exact keyword searches failing when users phrase questions differently (holiday/day off/vacation).

Click to watch →

Fabrizio Balliano: LLPhant: The PHP framework for generative AI @ PHPConPL24

Fabrizio Balliano: LLPhant: The PHP framework for generative AI @ PHPConPL24

466

This section explains embeddings and vector databases for semantic search. Embeddings are numerical vectors representing text that capture semantic relationships—similar concepts have similar vectors. The vectorization process transforms text into mathematical representations where operations like 'king - man + woman' yield vectors close to 'queen'. Generating embeddings involves: reading data, splitting into chunks (typically 500 characters), enriching with metadata, generating vectors (computationally intensive), and storing in vector databases. Supported vector stores include memory, filesystem, PostgreSQL, Chroma DB, Milvus, and Elasticsearch. This technology enables efficient similarity searches for retrieving relevant information from large document collections.

Click to watch →

Fresh AI Knowledge with Retrieval Augmented Generation

Fresh AI Knowledge with Retrieval Augmented Generation

795

Embeddings convert text into numerical vectors (typically 1,000-10,000 dimensions) that machine learning models can process. The encoding process transforms text through mathematical transformations, and decoding reverses this to retrieve original text. Vector databases store these embeddings rather than raw text, enabling semantic search that finds similar meaning rather than exact matches. This differs from traditional databases that search on exact values. Vector databases allow the system to find documents with similar semantic content even if exact words don't match, which is essential for understanding user queries.

Click to watch →

Deep Dive

The concept of Transfer Learning in machine learning, specifically how base models are adapted to specific downstream tasks.

02 ПЗАД: файнтьюн LLM в колабе

02 ПЗАД: файнтьюн LLM в колабе

396

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that reduces trainable parameters by using low-rank matrices. LoRA decomposes weight updates into two low-rank matrices A and B, where the rank R is much smaller than the original weight dimensions. This decomposition allows the model to learn task-specific adjustments while keeping the base model frozen. LoRA enables efficient fine-tuning of large language models by reducing trainable parameters from millions to thousands, allowing multiple task-specific adapters to be maintained for a single base model. Prompt tuning is a fine-tuning approach where additional parameters are learned to modify the input prompt rather than the model weights. The base model remains frozen, and only the prompt adapter is trained.

Click to watch →

State of GPT

State of GPT

3.3K

Pre-trained models learn powerful general representations that can be efficiently fine-tuned for downstream tasks. Around GPT-2, researchers discovered that models can be prompted effectively to perform tasks without fine-tuning. By arranging text to look like a document with a question, the model completes the document, effectively answering the question. This 'prompt engineering' approach allows base models to perform tasks by completing what they perceive as a document, rather than being explicitly trained for those tasks.

Click to watch →

AI Engineering: A Realistic Roadmap for Beginners

AI Engineering: A Realistic Roadmap for Beginners

81K

Prompt engineering involves creating effective prompts to reliably get LLMs to behave as intended. Fine-tuning involves taking a pre-trained foundation model and passing it additional data to adapt it for specific tasks. For example, using a smaller open-source model like Mistral and training it on 10,000 examples of hockey scores would enable it to make predictions based on that specialized data. Fine-tuning leverages the base knowledge already embedded in the model while biasing it toward specific tasks with limited training data.

Click to watch →

إعادة تدريب نماذج الذكاء الاصطناعي باستخدام لورا - Fine-tuning - LoRA

إعادة تدريب نماذج الذكاء الاصطناعي باستخدام لورا - Fine-tuning - LoRA

3.6K

LoRA fine-tuning allows models to adapt to specific tasks through prompt engineering. After fine-tuning, the model can be prompted with specific instructions to perform the desired task. For example, a model fine-tuned for quote classification can be prompted with 'Give me keywords for this quote' to extract thematic keywords. The quality of the fine-tuned model depends on the quality and quantity of training data, with more representative data leading to better performance. The model can generalize to similar inputs it hasn't seen during training.

Click to watch →

WeAreDevelopers LIVE - Machine Learning Day

WeAreDevelopers LIVE - Machine Learning Day

1.3K

Fine-tuning adapts pre-trained language models to specific tasks by continuing training on labeled data. For sentiment analysis, the model is trained to generate sentiment labels given input text. Encoder models (like BERT) process input and output contextual representations, while decoder models (like GPT) generate text autoregressively. Zero-shot learning uses pre-trained models without task-specific training by formulating tasks as completion problems. Prompt learning optimizes task descriptions automatically using a small neural network (prompt decoder) that generates prompt tokens, while the large language model remains frozen. This approach allows multiple tasks to be handled by the same model by specifying different prompt tokens.

Click to watch →

Deep Dive

Core principles of Prompt Engineering, including in-context learning and how system instructions guide model behavior.

What You MUST Know About AI Engineering | Chip Huyen, Author of “AI Engineering”

What You MUST Know About AI Engineering | Chip Huyen, Author of “AI Engineering”

59.3K

Prompt engineering requires systematic approaches despite seeming simple—version control, experiment tracking, and documentation are essential. In-context learning allows models to perform tasks by examining examples in the input context, eliminating retraining needs. Zero-shot learning works with no examples, while few-shot learning provides a few labeled examples. Defensive prompt engineering protects against harmful outputs through carefully crafted system prompts (behind-the-scenes instructions) and user prompts. Retrieval-Augmented Generation (RAG) remains valuable even with long context windows because models struggle to efficiently utilize extremely long contexts, and memory constraints will always limit processing capacity. AI judges (models evaluating other models) offer promising but require careful calibration to maintain consistent evaluation standards over time.

Click to watch →

Stanford Webinar - Large Language Models Get the Hype, but Compound Systems Are the Future of AI

Stanford Webinar - Large Language Models Get the Hype, but Compound Systems Are the Future of AI

140.7K

The GPT-3 paper established a general template for prompt design: context/instructions, followed by demonstrations showing desired behavior, then the target input. This pattern applies across tasks from question answering to translation to text completion. The model learns to imitate the behavior shown in demonstrations when given new targets.

Click to watch →

FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD

FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD

4.4K

Prompt engineering is the most important skill for engineers using AI agents. There are two ways to prompt agents: user prompts (single tasks) and system prompts (the law for every task). The system prompt is vastly more powerful because every word written is multiplied over every user prompt. A well-structured system prompt should include: (1) Purpose - defining the relationship as a no BS, clear, concise, actionable partnership focused on solving problems; (2) Instructions - concrete communication patterns. Positive patterns include: placing most important information last, using plain specific language, stating each fact once, matching detail level to task complexity, challenging incorrect assumptions directly, using simplest domain terminology, and optimizing for engineering value over quotability. Negative patterns to avoid include: loadbearing, worth stating plainly, here's the honest truth, carry the argument, overusing em-dashes, flattery without reason, decorative headings, and semicolons/fragments. Reference points are short codes assigned to sections of information (e.g., D1 for decisions, R1 for risks, FS for findings) that enable quick navigation. Hard operational boundaries prevent agents from doing unrequested work: deliver only what was requested, avoid cleanup/refactoring/documentation, not speculate on future requirements, not claim completion without evidence, and never add co-authors to commit messages. Aliases are short codes that expand into specific instructions: STR (simplify, compress, and repeat), ELI (explain like I'm 18), FOC (focus on what matters most), and REF (rewrite with reference points). In-context distillation involves providing concrete examples of desired and undesired responses to teach agents specific communication styles. The system prompt is the most effective place to prompt engineer because it's applied across every prompt going into the agent and every response coming out.

Click to watch →

Как сделать приложение на базе ChatGPT

Как сделать приложение на базе ChatGPT

6.9K

Prompt engineering involves reformulating queries to improve model responses. In-context learning is a powerful technique where users provide examples within their prompt to guide the model's behavior. By including few-shot examples of desired input-output pairs, users can significantly improve model performance on specific tasks. This works because LLMs can recognize patterns and apply them to new inputs, making it easier for the model to understand what the user expects.

Click to watch →

Instruction Fine-Tuning and In-Context Learning of LLM (w/ Symbols)

Instruction Fine-Tuning and In-Context Learning of LLM (w/ Symbols)

12.9K

In-context learning uses structured templates with instructions, exemplars, and evaluation examples. Researchers progressively simplified ICL prompts by removing instructions, replacing natural language labels with arbitrary symbols, and eliminating both. This revealed that larger models (540B parameters) can override internal semantic knowledge through ICL, while smaller models rely more on pre-trained priors. Flipped labels (contradictory examples) and semantically unrelated labels were tested to understand how models balance internal knowledge against presented examples. The number of examples per class critically affects performance: 2 examples yield ~50%, 4 examples reach ~70%, 8 examples achieve ~80%, and 16 examples approach ~90%. A fundamental distinction exists between ICL and fine-tuning: fine-tuning changes all attention weights and layers, fundamentally altering how the model learns. In contrast, ICL operates only on activations—manipulating hidden states during inference without changing parameters. Research revealed that ICL requires neural networks to implicitly construct a mapping from in-context examples to predictions without parameter updates, encoding an implicit context-dependent model in hidden activations.

Click to watch →

AI Concepts Overview

Simple AI Assistants

Fine-Tuning Models

RAG Architecture

AI Agent Basics

Choosing Approaches

Hybrid Systems

AI Concepts Overview

0:00

Playing Section

1
Video introduces key differences between AI models, RAG, and agents.
2
Aims to clarify when to use each approach or a hybrid solution.

Deep Dive

Hybrid LLM Architectures, such as combining Retrieval-Augmented Generation with Fine-Tuning (e.g., RAFT - Retrieval Augmented Fine-Tuning).

RAFT (RAG + Fine-tuning) in Azure AI Studio

RAFT (RAG + Fine-tuning) in Azure AI Studio

2.5K

RAFT (Retrieval-Augmented Fine-Tuning) is a hybrid approach that combines Retrieval-Augmented Generation (RAG) with fine-tuning to improve LLM performance on domain-specific tasks. Unlike traditional RAG which can retrieve irrelevant documents causing noise, RAFT generates synthetic training data by using a large language model to create questions and answers from domain documents, including oracle chunks (correct context) and distractor chunks (irrelevant context) to teach the model to focus on relevant information. This approach leverages chain-of-thought reasoning to maximize answer correctness while maintaining the efficiency benefits of fine-tuning smaller models. The method addresses the limitations of both pure RAG (irrelevant documents) and pure fine-tuning (high cost, limited flexibility) by creating a librarian-like system that knows which documents to consult for each question.

Click to watch →

LlamaIndex Webinar: Retrieval-Augmented Fine-Tuning (RAFT)

LlamaIndex Webinar: Retrieval-Augmented Fine-Tuning (RAFT)

7.1K

Retrieval-Augmented Fine Tuning (RAFT) is a technique that combines fine-tuning with Retrieval-Augmented Generation (RAG) by training language models on domain-specific datasets containing both relevant documents and distractor documents, enabling models to learn how to selectively focus on relevant information while ignoring irrelevant content during inference, thereby achieving superior performance on open-book exams in specific domains compared to either fine-tuning alone or naive RAG approaches.

Click to watch →

生成系 AI のファインチューニングと Amazon SageMaker JumpStart | AWS AI Week For Developers (アドバンストトラック)

生成系 AI のファインチューニングと Amazon SageMaker JumpStart | AWS AI Week For Developers (アドバンストトラック)

4.2K

RAG and fine-tuning can be combined in hybrid architectures. The system first checks if the query topic exists in the fine-tuned model's training data; if so, the fine-tuned model handles it directly. If not, RAG retrieves external information and the fine-tuned model generates the response. Model selection should consider: (1) external knowledge injection needs, (2) model behavior change requirements, (3) cost optimization needs, and (4) implementation complexity. RAG is easier to implement, while fine-tuning requires more effort but offers better cost efficiency for specific tasks.

Click to watch →

Основы RAG и Fine-Tuning. Учим приложение отвечать на вопросы в контексте ваших документов

Основы RAG и Fine-Tuning. Учим приложение отвечать на вопросы в контексте ваших документов

272

Large Language Models predict the next token based on input sequences, trained on massive text corpora. They hallucinate because they always generate something, even when uncertain. Retrieval Augmented Generation (RAG) addresses this by providing additional context documents alongside queries. Instead of retraining the model, RAG enriches input with vector representations of relevant documents. This allows rapid knowledge updates without model retraining. The architecture involves: (1) converting documents to vectors using embedding models, (2) chunking documents with overlap to preserve semantics, (3) storing in vector databases like FAISS, (4) converting queries to vectors, and (5) retrieving relevant chunks for the LLM.

Click to watch →

Create Dataset Locally with Retrieval Aware Fine-tuning (RAFT)

Create Dataset Locally with Retrieval Aware Fine-tuning (RAFT)

4.8K

Retrieval Aware Fine-tuning (RAFT) is a Microsoft-developed framework that bridges the gap between traditional fine-tuning and Retrieval-Augmented Generation (RAG) by creating synthetic datasets where models pre-study documents before retrieval, thereby improving answer quality by reducing the risk of retrieving irrelevant or hallucinated documents that plague standard RAG systems.

Click to watch →

Deep Dive

Hands-on implementation of autonomous workflows using agentic frameworks like LangChain, CrewAI, or AutoGen.

Modern AI Agents from A Z Building Agentic AI to Perform Com with Sinan Ozdemir

Modern AI Agents from A Z Building Agentic AI to Perform Com with Sinan Ozdemir

772

Several frameworks implement agentic workflows: LangChain provides Pythonic ML components with LangGraph for stateful multi-actor applications, offering structured output constraints. CrewAI is popular but may not consistently complete tasks reliably. Autogen by Microsoft focuses on distributed real-time applications across multiple languages. Swarm is experimental from OpenAI focusing on agent handoffs. The speaker recommends writing custom agent prompts when specific requirements don't match existing frameworks, noting that replicating core features of frameworks like CrewAI requires relatively little code. Practical implementation involves setting up virtual environments, importing tools, and instantiating agents with roles, goals, and backstories. Code examples demonstrate how agents follow the React paradigm, selecting tools, executing them, and synthesizing results into final responses.

Click to watch →

Andrew Ng's 2-hour course covering graph engineering. 1 prompt → 100 agents → loops → graphs

Andrew Ng's 2-hour course covering graph engineering. 1 prompt → 100 agents → loops → graphs

5.5K

Agentic workflows represent a paradigm shift from traditional LLM usage, employing iterative approaches where AI can write outlines, conduct research, draft content, critique it, and improve it. This produces better results than one-shot generation and contributes to training foundation models. Langchain provides comprehensive support through prompt templates, tools, and memory. The framework supports React agents (reasoning and action), self-refinement agents, and AlphaCode-style coding agents. Langgraph extends Langchain for agent flows with nodes, edges, and conditional edges. State is tracked throughout execution and can be persisted for resumption. State variables can be annotated with operators like 'add' to control how updates are handled.

Click to watch →

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

478

Agentic AI frameworks are toolkits for building systems that perform planning, acting, and iterating, unlike simple chatbots. These frameworks provide predefined architectures, integration tools, monitoring capabilities, task management, and communication protocols. Most agentic AI systems fall into five categories: (1) Linear workflows for predictable step-by-step processes like customer support agents, using LangChain, LlamaIndex, or LangGraph; (2) Autonomous agentic systems for open-ended problems where agents collaborate independently, using AutoGen, Baby AGI, or CrewAI; (3) Role-based systems with defined agent roles and constraints, using CrewAI or AutoGen; (4) Production orchestration systems for real-world business integration, using Agent Framework or LangGraph; (5) Rapid prototyping tools for quick idea validation, using LangFlow or Flowise. The key to framework selection is matching the system type to your needs: use workflow approaches for predictable systems, autonomous agents for exploratory problems, role-based systems for teamwork, production frameworks for deployment, and prototyping tools for testing ideas.

Click to watch →

Agentic AI with Langchain , Llmaindex , Langgraph & Crew AI | Live Lecture | Euron

Agentic AI with Langchain , Llmaindex , Langgraph & Crew AI | Live Lecture | Euron

1.4K

Agentic AI frameworks enable autonomous AI agents. The four main frameworks are LangChain (foundational), LangGraph (complex multi-agent workflows), LlamaIndex (data indexing and RAG), and CrewAI (role-based multi-agent systems). An AI agent is a system capable of autonomous decision-making, unlike traditional automation with rigid if-then rules. Tools are callable functions agents invoke to perform tasks. LangChain enables creating agents through a systematic process: tools are Python functions decorated with @tool that agents can call, such as math calculators using eval() or poem generators using LLMs. A React agent uses the thought-action-observation pattern, requiring an LLM, tools, prompt template, and output parser. The prompt defines the agent's role, available tools, and expected output format. Execution uses an AgentExecutor that runs the thought-action-observation loop, with verbose mode showing the agent's reasoning process. The agent automatically decides which tools to call and in what sequence based on user requests.

Click to watch →

Best Frameworks To Learn For Building AI Agents And Agentic AI

Best Frameworks To Learn For Building AI Agents And Agentic AI

77.2K

LangChain provides a comprehensive framework for building generative AI applications with extensive integrations for major LLM providers (Google Vertex AI, OpenAI, AWS, Anthropic), document loaders for various formats, and vector store connections. LangGraph extends this by introducing graph-based architectures using directed/cyclic graphs and stateful nodes, enabling autonomous AI agents to execute complex workflows without human intervention beyond initial queries. Together, these frameworks support granular workflow modifications, RAG implementations, and independent agent creation capable of making autonomous decisions to solve complex tasks.

Click to watch →

Deep Dive

Evaluation methodologies for advanced LLM systems, including frameworks like Ragas for RAG and validation techniques for agentic trajectories.

Тихомиров М.М. - 4. BERT, Sentence BERT, векторные БД, RAG

Тихомиров М.М. - 4. BERT, Sentence BERT, векторные БД, RAG

145

Ragas is a framework for evaluating RAG systems. It evaluates both retrieval quality (how relevant retrieved documents are to the query) and generation quality (how accurate and relevant the generated response is). Ragas uses LLM-based evaluation (asking another LLM to score responses), which has limitations (language bias, cost, potential hallucinations) but provides a practical approach to RAG evaluation.

Click to watch →

A Survey of Techniques for Maximizing LLM Performance

A Survey of Techniques for Maximizing LLM Performance

220K

The RAGAS framework evaluates RAG systems across four metrics: (1) Faithfulness - checks if answer facts can be reconciled with retrieved content, identifying hallucinations; (2) Answer Relevancy - measures whether the model uses retrieved content appropriately relative to the original question; (3) Context Precision - evaluates the signal-to-noise ratio of retrieved content by determining which content blocks were actually used in the answer; (4) Context Recall - measures whether the relevant information needed to answer the question exists in the content. These metrics help identify whether to optimize retrieval, add reranking, or adjust embeddings.

Click to watch →

【2024/06/06】【LT大会#7】LLMの活用・機械学習・データ分析関係のいろいろな話題にふれよう【アーカイブ】

【2024/06/06】【LT大会#7】LLMの活用・機械学習・データ分析関係のいろいろな話題にふれよう【アーカイブ】

902

This presentation introduces Ragas, an open-source evaluation framework for Retrieval-Augmented Generation (RAG) systems. Key concepts include: (1) Ragas as a GitHub-hosted open-source project with accompanying academic papers, (2) The distinction between offline evaluation (using prepared datasets) and online evaluation (using real user feedback), (3) Three evaluation categories: retrieval component evaluation, generation component evaluation, and end-to-end evaluation, (4) Evaluation metrics implemented using LLM-based assessment and embedding-based similarity measures, (5) The framework's structure with MetricWithLLM and MetricWithEmbedding classes, (6) The ability to integrate with LangSmith for evaluation, (7) Ragas's synthetic data generation capabilities for creating evaluation datasets, (8) The importance of understanding evaluation metrics for improving RAG system performance.

Click to watch →

RAGの評価フレームワーク ”RAGAS”を徹底解説

RAGの評価フレームワーク ”RAGAS”を徹底解説

1.8K

RAGAS is an automated evaluation framework for Retrieval-Augmented Generation (RAG) systems that uses large language models to assess the quality and relevance of generated responses through multiple metrics including faithfulness (accuracy based on retrieved context), answer relevance, context precision, context recall, entity recall, semantic similarity, faithfulness, side critique, and summary score, enabling efficient and objective evaluation of RAG system performance without relying on human assessment.

Click to watch →

Want to Master Gen AI Models? Watch This RAGAs Evaluation Now | RAGAs Framework | Satyajit Pattnaik

Want to Master Gen AI Models? Watch This RAGAs Evaluation Now | RAGAs Framework | Satyajit Pattnaik

10.2K

RAGAS provides five specialized metrics for comprehensive RAG pipeline evaluation: Answer Correctness measures factual accuracy against ground truth; Answer Similarity assesses semantic equivalence; Context Recall evaluates how well retrieved documents support answers; Faithfulness checks if answers accurately reflect retrieved context without fabrication; Context Precision measures relevance of retrieved documents to questions. Practical implementation involves building a complete RAG architecture with document loading, chunking, embedding, vector database storage, retriever creation, and prompt engineering. Evaluation data preparation includes defining questions, preparing ground truth answers, generating model responses, and collecting retrieved contexts. Converting this data to structured format enables RAGAS evaluation by importing metrics and calling the evaluation function with the dataset, metrics, LLM, and embeddings to obtain quantitative performance scores. Interpreting results guides targeted optimization of retrieval or generation components based on which metrics require improvement.

Click to watch →

Deep Dive

Cost, latency, and deployment optimization strategies for serving and scaling production-grade LLM applications.

Enabling Cost-Efficient LLM Serving with Ray Serve

Enabling Cost-Efficient LLM Serving with Ray Serve

8.6K

LLM serving is expensive due to high GPU costs ($3,000/month per GPU). Cost efficiency requires optimizations across infrastructure, serving, and model layers. Infrastructure optimizations include: (1) Auto scaling automatically scales each model up and down based on traffic, achieving 50% cost savings. (2) Heterogeneous compute uses smaller instances for smaller models and larger instances for larger models, with hybrid cloud support across AWS, Oracle, and GCP. Model-level optimizations include: (1) Continuous batching, which creates new batches after each forward pass to handle varying sequence lengths, achieving 5.5x throughput improvement over optimized static batching and 23x over naive approaches. (2) Speculative decoding, which uses a small draft model to predict K tokens ahead with large model verification, reducing per-token latency by up to 50%. Ray LLM packages these optimizations into an easy-to-deploy solution supporting any Hugging Face model.

Click to watch →

Mastering LLM Inference Optimization From Theory to Cost Effective Deployment: Mark Moyou

Mastering LLM Inference Optimization From Theory to Cost Effective Deployment: Mark Moyou

29.7K

LLM inference operates fundamentally differently from standard deep learning inference, generating one token at a time while maintaining the entire conversation history in a Key-Value (KV) cache on the GPU, which directly impacts deployment cost and performance; effective optimization requires understanding four distinct querying patterns (long input/short output, long input/long output, short input/long output, short input/short output), measuring key metrics like time-to-first-token and inter-token latency, and leveraging techniques such as quantization, tensor parallelism, and in-flight batching to maximize throughput while minimizing costs.

Click to watch →

LLM Explained

LLM Explained

109.3K

Five techniques separate demo builders from production engineers: (1) Retrieval-Augmented Generation (RAG) embeds proprietary documents into vector databases and retrieves relevant chunks at query time, enabling accurate answers about private codebases. (2) Prompt caching reduces latency by up to 80% when system messages are identical across requests. (3) Token budget management counts tokens before every API call and trims oldest non-system messages when exceeding 80% of context window. (4) Model routing routes simple tasks to small fast models (GPT-4o-mini) while reserving GPT-5 for complex reasoning, cutting costs by 60%. (5) Eval-driven iteration runs curated test suites (50-100 prompt-response pairs) before deployment using tools like Braintrust and LangSmith.

Click to watch →

How to Evaluate LangChain Pipelines: Quality, Cost, and Latency Optimization

How to Evaluate LangChain Pipelines: Quality, Cost, and Latency Optimization

101

Systematic evaluation of LangChain pipelines is essential for production-grade AI applications, requiring developers to balance three interconnected metrics—quality (accuracy and relevance of responses), cost (API expenses from LLM providers), and latency (response speed)—through offline evaluation strategies, A/B testing different models and chain variants, and cost optimization techniques such as prompt compression, model cascading, caching, and batch processing, all supported by automated testing and continuous monitoring to make data-driven decisions without sacrificing user experience.

Click to watch →

Shared vs Private LLMs: Cut Latency, Costs & Gain Control | Predibase Inference Engine Deep Dive

Shared vs Private LLMs: Cut Latency, Costs & Gain Control | Predibase Inference Engine Deep Dive

216

LLM deployment economics involve strategic trade-offs between hardware costs, throughput requirements, and latency constraints. Dedicated open source deployments can achieve half the cost of equivalent commercial shared endpoints at scale, with FP8 quantization providing 2x throughput improvement and speculation (turbo) enabling 3x gains by predicting multiple tokens simultaneously. Combined, FP8 and speculation deliver up to 4x throughput improvement over base FP16 models with minimal overhead. Data privacy strategies progress from standard dedicated deployments to VPCs on customer-owned hardware, culminating in direct ingress eliminating all intermediary routing between applications and LLMs. Network latency reductions range from 10ms (collocated regions) to 220ms (cross-continent) depending on geographic separation from control planes. Advanced production deployments leverage dynamic LoRA adapter switching enabling hundreds of specialized variants on single deployments with minimal throughput impact, combined with hardware selection flexibility from A10G to H100 GPUs and FP8 quantization options for optimal cost-performance trade-offs.

Click to watch →

AI Concepts Overview

Simple AI Assistants

Fine-Tuning Models

RAG Architecture

AI Agent Basics

Choosing Approaches

Hybrid Systems

AI Concepts Overview

0:00

Playing Section

1
Video introduces key differences between AI models, RAG, and agents.
2
Aims to clarify when to use each approach or a hybrid solution.

4.1K views12827:46@sunnysavita10Original Release: 2025-05-27

This video explains three key AI approaches: (1) Simple AI assistants use pre-trained LLMs for basic text generation tasks; (2) Fine-tuning retrains a pre-trained model on specific domain data to create expertise in particular topics; (3) RAG (Retrieval-Augmented Generation) connects LLMs with external data sources like vector databases to provide up-to-date knowledge and context; (4) AI agents combine LLMs with tools and actions, enabling autonomous systems that can think, observe, and perform tasks. These approaches can be combined in hybrid architectures—for example, fine-tuning a model on domain data and then building RAG or agent systems on top. The choice depends on requirements: use simple LLMs for basic chat, fine-tuning for domain expertise, RAG for real-time knowledge access, and agents for autonomous task execution.

[00:00:00]Hey. Hi everyone. Welcome back to my YouTube channel. My name is Sunonny Savvita and I'm back with another exciting and important video. So guys, in this particular video, we're going to understand the differences between the simple AI assistant, your fine-tuned model, rag and the AI agent. So guys this was the very mandatory video and intentionally I kept the video so your each and every concept could be clarified with respect to this fine-tuning rag AI agents and easily you can understand the differences between each other because uh so many people are having a like doubt that which one should I use should I go with a finetuning rag or AI agents or should I use the hybrid approach so inside this particular video I'm going to discuss all this thing and your each and every doubt is going to be clarified. So please make sure guys you are going to watch this video till the end because at the end I'll tell you which one you need to use when and I will uh show you some uh uh fine-tuned model so that directly you can use those particular model inside your rag architecture or inside your agents. Uh so guys as you know I started with this finetuning series and I already recorded four video inside this one. So in the first video I discussed the syllabus in the second I discussed the complete introduction of the finetuning in the next video I discuss about the transfer learning and then in the last video I discuss about all the framework and all the resources regarding the finetuning. Now in this particular video we'll see the differences and right after this one guys we going to jump to the practical.

[00:01:41]Yes. So I'll show you the practical of each and every technique of the finetuning in my YouTube playlist itself. So if you haven't subscribed so far guys, please subscribe the channel and hit the bell icon. So whenever I'm uploading a video, you will get a notification. Now without wasting a time, uh let's start with the video guys. So here guys, I kept all the points which I am going to discuss. At the first place, we'll talk about the simple AI assistant. uh then we'll see about the then we'll see the definition of the fine tuning then we'll talk about the rag and then I'll come to the AI agents so each and every architecture I kept over here and then the differences between each one of them uh and then we'll talk about which one we should use when so here also I kept some architecture and some real time analogy so uh let's start with this AI assistant itself and I hope so you understood about the simple Simple AI assistant. So what is a simple AI assistant guys? And one more thing uh let me clarify over here. So whatever architecture I'm going to show you right whether it's a simple AI assistant or whether it's a LLM fine tuning or it's a rag or it's a agent. So one thing would be in a center and that particular thing is going to be large language model. Got it? So guys this large language models are nothing this large language model is a AI model is a deep learning based model which was trained on a very huge amount of data right so this large language model is trained on a huge amount of data which is having millions and the billions of parameter. So if I'm talking about the parameter I think you know the meaning of the parameter if you don't know you will get to know in my upcoming videos where I will discuss about the Lora and Qura. Uh parameter is nothing parameter means weight and biases. If you have a basic understanding of the deep learning or machine learning you can easily understand the meaning of this particular term. Uh now guys LLM are nothing LLM are the model which was trained on a very huge amount of data.

[00:04:01]If you don't know about it, you can look into my previous video that I already discussed. Uh so this LLM is capable to answer any sort of a question. This LLM is capable to writing any sort of a paragraph, essay, translating something, writing code, chatting, summarizing. So in short, I can give you the example of the chat GPT. Chat GPT have been built uh on top of one LLM. The LLM is called the GPT. GPT uh there is a different different variants of the GPT which you will get now. So GPT 404 mini 03 right these are some advanced GPT based model and this was uh this was used inside the chat GPD application and it is capable of each and every for each and every task. Uh now guys whenever we talking about the simple AI assistant so simple AI assistant means what? So we have one LLM which we have trained on a very huge amount of data and to that particular LLM we are passing input and we are generating a output and if we are going if we are going to build any sort of a application using this particular LLM right uh so those uh that particular application is called simple AI assistant as you know uh in my previous video many time I explain you this particular thing so guys uh again I'm repeating so we have a llm to this LLM we are passing input and we are generating output. Now using this particular LLM we are going to be create one simple chatbot or one simple QA system. So that is nothing that is called simple AI assistant. Now this LLM could be uh my raw model. So what's the meaning of the raw model? Raw model is nothing. It's a unsupervised model.

[00:05:39]Okay.

[00:05:40]Unsupervised pre-trained model. So unsupervised pre-trained model means what? So if we're talking about the unsupervised pre-trained model uh so the meaning is it it was trained on a very huge amount of internet data okay huge amount of internet data and the data was not labeled. So for uh training this particular model we have used something or something which is called auto reggressive training. Auto reggressive training. In this auto reggressive training we always predict the next word. Okay. What we do guys in the auto reggressive training we always predict the next word inside the given sentence and this thing I explained you in my previous video if you don't know about it you can go and check it out so auto reggressive training and in that we are going to predict the next word and uh we are going to be train our model so this unsupervised pre-training is also called self-supervised learning in my previous video I clarify all this thing if you don't know about it you can go and check with that but with that particular video. Uh now guys uh this is clear. I hope we have a simple LLM. We have trained it on a huge amount of data and with that we are going to be generate an input. We we are going to be generate some output based on the given input.

[00:06:55]Now coming to the next topic over here.

[00:06:57]The next topic is a finetuning. So further what we can do we can fine-tune this particular model means we can fine-tune this model on some specific task. Uh so what does it mean? What is the meaning of that specific task? So the task could be anything. Let's say uh I want to fine-tune that on my uh own company data on my own domain specific data for what for the uh medical for the medical chatbot or what we can do so uh I have a company let's say I have a fintech company and in fintech company the requirement would be different compared to this medical company okay so what I will do here again I will collect the fintech data and again What I will do again I will train my model again I will retrain my model which retrained model which uh pre-trained model the raw model right the foundation model or which is also called the pre-trained model right so we are going to be fine tune or we are going to be retrain that pre-trained model retrain the pre-trained model okay and that pre-trained model is also called the raw model or the foundation model so we could have a multiple example and this already I discussed in my previous video if you will go and check there you will find out in a very detailed Okay. So let's uh read the definition. So what I written over here.

[00:08:13]So here I written what we are doing we are just taking a pre-trained model and which already train on a huge amount of data set and then what we are doing we are going to be retrained again on some specific data set. Uh so here you can see imagine you have a you you hired a person who already know the English.

[00:08:31]This is just a example. This is just an analogy uh which I mentioned over here.

[00:08:36]So let's say you have a like person who you you hired it for the English, Hindi and for the basic science but you want them to expert in a medical science. So what you will do you will again uh train that particular person on some medical related information which is already aware about the English, Hindi and the basic science. Right? So finetuning means again we are going to be retrain the model. Again we are going to be retrain the model on some specific data.

[00:09:00]So this architecture will clarify everything. So here we have a base model here you can see. Then we are going to be fine tune on some specific data. And now what we got? We got some custom model. So I think I no need to spend much time on this finetuning. Now let's uh look into the next architecture which is a rag one. Guys, this is the architecture of the rag. If you have seen my YouTube channel, uh I think I explained this particular architecture maybe 100 times.

[00:09:29]I uh I recorded one very comprehensive playlist on my YouTube channel. uh if you don't know about it you can go and check you will learn a lot regarding this rag R A okay so in the RA what we have see guys u I told you one thing you know it's starting itself so whenever we are talking about any of the concept so the LLM will be the center of it okay so here also you can see the LLM this one now uh whenever we talking about let me highlight the LLM I think you can see this is the LLM guys whenever we talking about the rank so in the rag actually we are going to be connect our LLM with this here you can see this external data source okay we have this external data source or this could be anything it could be any database this could be any API this could be any web page this could be any document so whatever you can think with respect to the data this uh could be anything what we are doing we are going to be store this particular data inside the database Now this database generally is called a vector database because we are going to be vectorized the information. Now why we are doing that vectorization so that we can perform the retrieval on top of it.

[00:10:44]Okay. So here is the retrieval pipeline this one. So we pass the query to the vector database and we fetch the relevant data. Okay. Now why we are fetching this relevant data? So that we can provide that relevant data or we can provide the appropriate context to the LLM. Okay. So the idea of the rag idea behind the rag is very simple. Uh we can connect our data external data. So you can say this is nothing. This is our external data to lm that's it. And we are storing it after doing a chunking embedding right where we are storing in the database.

[00:11:27]And then we are going to perform the retrieval. So why we are doing it? so that we can provide this particular data the right data the contextual data and the data based on the user query to the LLM itself that's it right so that's a simple meaning of rag I think you already know so let me revise all the things so we're talking okay I I'll revise that let me complete the AI agent also and after that I'll revise it guys when we're talking about the rag so rag is stand for retrieval argument generation and here we have a retrieval we are going to be retrieve the relevant document we are passing it the llm and then we are passing it to the sorry then we are going to be generate a final answer. So we have a user question. We are retrieving the information relevant document. Then we are passing to the LLM and then final answer. And from where we are going to be retrieved the information we are going to be retrieve the information from the vector database. If you have any sort of a doubt regarding the rack, you can look into this particular architecture. It will clarify your doubt in a better way.

[00:12:22]Okay. Now there is some more information regarding the rack. So let's say we want to connect our LLM with some realtime info. Okay. Realtime info. uh let's say my LLM was trained last year last year. Okay. And now this year I'm using this particular LLM and and and I'm asking who won the world cup who won the World Cup in 25. So my LLM will not be able to answer. So in that case what I will do I'll connect this particular LLM with some external data sources. Okay. Here I have stored some data which I fetch maybe from some web pages or maybe from some API the realtime data and then I'm providing this particular data to the LLM some context okay and then we are generating a final answer using this LLM. So why we use this rack guys to connect our LLM with a realtime information okay to uptodate for up to for updating my LLM knowledge in the form of context and one advice I can give you with respect to this rag so whenever we are talking about the rag guys okay so guys always try to use rag with some good reasoning based model in some whenever you are going to build a rag based architecture now so use some good reasoning model. Now this reasoning model will be capable to understand your complete context. Okay. So this is the one of the advice which I wanted to give you. Now let's move to the next one which is AI agent. Now guys uh whenever we are talking about the AI agent, so we should always keep couple of thing in our mind.

[00:14:11]The first one uh in the AI agents okay the AI agents it's having a capability of the thinking it's having a capability of taking a action it's having a capability of making a observation now how so guys uh whenever we are talking about the AI agents now you need to keep one more thing in your mind which thing if you have attended my previous uh session if you have seen my previous video then you can easily answer. First is a LLM and the second is a tools. Okay. So this LLM is nothing.

[00:14:47]This LLM actually it is a brain of the agent and this tool is nothing. It's a action. I can give you very good example. Let's say uh human human is having brain. Okay. And human is having the other part of the body, right? Let's say hands and legs. Now brain is giving a brain is giving a instruction. Okay, a brain is giving a instruction to the hand. Okay, pick something. Let's say pick this mouse. Pick this phone. Yes, I'm able to pick it, right? Let's say my brain is giving a signal to my uh legs. Okay, run, walk or stop. So, I'm able to do that particular thing. So here in the same way LLM is a brain okay that is a artificial brain which is giving a instruction to the tool okay and the tool is nothing it's called the actions getting my point so in a similar way see brain human brain it's giving a instruction okay and based on this particular based on the given instruction my hands and legs is taking a action now you can again compare so lm is nothing it's a brain it's giving some sort of a like a instruction right and based on that particular instruction we are taking a action how we are taking action via this tools so this hand and legs is nothing it's my tools okay this action is being generated by the brain okay and here the action is being generated by the llm I can give you a very good example let's say here uh uh you written a mail okay so let's say you return one mail now this mail actually you can write using the llm But if you want to send this particular mail, okay, whatever mail you have written, if you want to send it to someone, you cannot send using the LLM.

[00:16:46]For that, you required some external API, right? So you required some external API, some external code, right? So this API, this functionality, it is a tool. Okay? Using this particular tool, using this particular API, you are performing a action. So LLM is your brain which is capable to write a mail but API and the tool okay which is performing a action for you for what for sending a mail right so that is what that is a tool right using that particular using the LLM what you are taking you are taking some instruction that instruction is nothing it is coming in the form of prompts okay you are passing a prompt you are taking that action and you are calling this appropriate tool for sending a mail that's it guys now uh here I have dro draw an uh diagram. So here see uh we have a brain in between.

[00:17:37]So here we are giving an instruction we have a tool calling capability okay we have we can sustain the memory okay we can do the planning okay based on the multiple iteration and observation and all everything we can keep and here we have output so in short I can say uh again I'm repeating so in short I can say AI agents are nothing it's a smart AI assistant it's a smart AI assistant Okay, AI agents is nothing. It's a llm plus action. That's it guys. Okay. Uh whenever you are not able to think about this LLM and action, you can recall this particular example writing a mail and sending a mail. Guys, uh I hope we understood about the AI agents rag and all everything individually. Now let's look into the different and see uh which one we can use when. So what is the difference? I think now you easily you can uh uh write a differences. So LLM the normal LLM is nothing just a pre-trained model.

[00:18:46]Finetuning is what a mega LLM expert in a specific topic. Okay with some uh data we are retraining the model. Here retraining would be included. Now we're talking about the rack. So in the rack what we are doing we are going to be connect our LM with some external data source.

[00:19:11]Okay. Now here uh this external data source EDS is nothing. It is called the knowledge base also. It is a database. Got it? Now if you're talking about the agents guys. So as agent is llm plus tools. Now this tool could be anything.

[00:19:30]Anything means anything. Okay. It could be any realtime API. It could be any database connection. Okay. It could be it could be any database connection. It could be uh maybe whatever right?

[00:19:44]Whatever uh logic and the functionality you want to write it could be anything.

[00:19:48]And in LLM there would be one more thing. It is capable to think.

[00:19:53]It's capable to make a observation and how based on the prompts based on the given instruction.

[00:20:03]So agent rag is a different where we are just going having a database vector database but agents having some intelligence thinking capability observation capability based on the given instruction and the tool calling capability and this tool calling could be anything not only the vector not only any database it could be any database any API any sort of a custom logic okay so here is again one more level of definition this is just for you if you want to make a more clarity on top of this particular topic you can do Now uh the next question which one we should use when I think you are waiting uh for this particular question only now uh let me clarify guys when we should use when which one and one more very important question is there uh can we use together can we use some hybrid approach so let me tell you all this thing with my experience what I what I feel uh because I work in a real time I did fine tuning rag everything and what I feel guys for sure I'll I I'll discuss that. So first let's look at the formal thing formal uh uses of it. So we have a normal AI assistant right normal LLM model. So we we can use for what for doing a chat text generation chatbot writing answering and all everything I told you.

[00:21:20]Now let's say we have to train that particular model on some specific data.

[00:21:24]Okay to make it export. So here we can fine-tune it on that specific data. Now let's say we want to connect my model with some uh knowledge base with some for some live knowledge okay we can use the rag architecture where uh data injection retrieval uh would be included now if you want to build a autonomous system uh so there we want to give the capability of the thinking search plan and all so we we are going to build the AI agents right now in the starting itself I told you whatever architecture we are going to build the one thing is going to be common that's going to be this LLM so here is also LLM Here also this LLM this one. Here also this LLM. Okay. And here also this LLM this one. Right. So the my main question is uh okay individually we can do anything mean we can do the finetuning and we can use that model. We can create a rag right and yes the rag I already told you use the good reasoning based model. Rag will work in a better way. Same goes with the AI agent also use some good model and for sure it will work in a well manner. Uh right. But guys the question is can we use both all together? Can we combine this particular thing? Yes or no. So for that also here I written some sort of a thing with a real life analogy and for sure you will be able to understand. So guys here you can see uh if we we have to do some very basic task. Okay. So in that case do use the simple LLM without like fine tuning and all if it is not required. If finetuning is required, you can fine-tune it and in straightforward manner you can do that rag. I already told you when you have to use whenever you have to connect your LLM with a with some external data sources with some live knowledge and all you can create a rag and if you want some autonom autonomy inside your application use the uh agents okay now the thing is basically uh let's say we have llm we have one llm now this llm okay so we have fine- tuned this llm on some domain specific data fine-tune on this particular LLM some domain specific data. Now, cannot we use this fine-tune model for further for the rag architecture? Yes, we can use it. Who is stopping to us? Cannot we use this fine-tuning model for a for creating our agent application, agentic based application? Yes, we can do it. Who is stopping us? So, I hope you understood now. So yes the combined solution absolutely the combined solution is possible and that could be my best system. So first we have a base model the foundation model we can fine-tune it on some specific data set and on top of it either we can build a rag or agent and nowadays guys some agentic rag is also possible. I already taught you on my YouTube channel you can go and check out that. So here you can understand with some real life analogy. So let's say uh you have a brain you have a like train your brain okay you have fine tuned the brain uh means you got a expert brain right on some knowledge and all maybe some uh someone is teaching you something or you are observing from the uh surrounding uh then basically you are going through the internet and then you are going to be retraining your brain right means sorry you are going to be connect your brain with some Google knowledge Wikipedia and books and all okay and now see you have a brain, you expertise your brain, you connect it with some uh external data sources and then agents what you are doing. So you have a expert brain with some external knowledge uh external knowledge which you connected with the internet and like which you got which sorry which you train yourself okay with your surrounding knowledge and then you connected yourself with the internet and all and then on top of it on top of it you having a capability to do something extra right with your actions. So now you have all the knowledge with that you can explore the phones, tablet, car, bike or you can do any sort of a thing.

[00:25:24]So this is just a real life analogy which came in my mind that's why I kept it over here. You can think in a more better way or you can just look into this particular uh like example. So you have fine to the model just to change your tone tone according to the domain.

[00:25:37]Okay. Uh on top of it you can create a rag layer. Okay. And then the same rag right. So uh not not the normal rag you can create a autonomous system some advanced rag that's called the agentic rag. So this uh kind of solution is possible. I hope you understood. Now what I'm doing guys I'm showing you some good model uh from the internet itself.

[00:25:59]So here uh here I find out some fine-tuned model and your task would be over here. So you need to download this finetune model. See here is a DC coder 676 7B inst. So this model is specifically was uh trained on the coding data set and specifically for the cod coder and all. You can utilize this particular model and on top of it you can create your rag application. Okay where you can uh connect your this particular model with some external github repositories and repositories and all. Now on top of it itself you can create a agentic flow right you can connect your uh uh this particular model using some tools and al some real API for fetching a code or anything uh and then based on that you can generate a answer now here is one of the model from the mist only so open homes and this model have been trained on a huge amount of data guys with the mythology homes design and all. So you can utilize this model and on top of it you can create your own rag architecture and in the going forward classes we are going to retrain our own model for sure I'll show you that from a scratch. Now here is one more model. So this model actually it's a llama model again it was finetune on some external data. So you can identify on which data it it have been trained and then on top of it you can create a rag or agent or agentic rag anything. So on top of this fine-tuning model agent rag everything is possible and we can do this thing in a custom way and for sure we are going to do it in a upcoming session. So yeah this is it for this particular video. I hope your each and every doubt is clear now and you can explore more about it from your end as well and if you have any doubt you can ask me in a comment section. So until guys uh thank you. Bye-bye. Take care.

[00:27:42]I'll see you in the next video.

Up Next

Direct Preference Optimization (DPO): Fine-Tune LLMs Without RL

@SerranoAcademy

29.4K views•2024-06-21

Secure Multiparty Computation (MPC): Foundations & Challenges

@SimonsInstitute

7.3K views•2015-05-28

Bypassing Tor Censorship: Bridges and Pluggable Transport Guide

@Coding_ForEveryone

397 views•2024-06-11

Neural Networks Explained: Math, Layers, and Learning Fundamentals

@3blue1brown

21.9M views•2017-10-05

Related Study Plans & Knowledge Roadmaps

Structured learning paths in Artificial Intelligence

View All Artificial Intelligence Topics

artificial-intelligence

Aligning LLMs: DPO, TRL & Custom Datasets

01Foundations of Python & Machine Learning

02Transformers and Large Language Models (LLMs)

03Aligning LLMs: SFT and RLHF

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

Federated Learning: PyTorch, Flower & Privacy

01Python & Machine Learning Foundations

02Deep Learning Basics with PyTorch

03Introduction to Federated Learning

+ 2 more learning modules

Explore Roadmap

artificial-intelligence

Securing LLMs: Guardrails AI & NeMo Gateway

01Introduction to LLMs, Python, and APIs

02Understanding LLM Vulnerabilities & Threats

03Defending LLMs with Guardrails AI

+ 2 more learning modules

Explore Roadmap

artificial-intelligence

Building Recommenders: Two-Tower & Vector DBs

01Foundations of Python & Machine Learning Basics

02Introduction to Recommender Systems

03Deep Learning & Representation Learning

+ 3 more learning modules

Explore Roadmap