Deep Agent - Multi Agent RAG with Gemini and Langchain

Explore how ai agents are built with lm memory tool and planning tools, guided by a pocket dictionary mind map that outlines agent patterns and when to use each pattern.

An AI agent autonomously perceives with sensors, reasons in Elm, and acts through tools and memory to solve problems.

Identify four major agent patterns: react agent, plan and execute agent, multi-tool agent, and self-corrective agent. Learn how these patterns plan, act, verify, and adapt for real projects.

Compare Google Gemini models and pricing to choose the right model. Learn about context window, token costs, thinking levels, and grounding with Google search in Python and Langchain.

Generate a Lang Smith API key, configure env variables, and verify loading in the notebook for agent tracing via the Lang Smith dashboard.

Test your Google Gemini API key using the sample code from Google AI Studio, then validate Lang Smith integration and Lang chain setup for multi-agent RAG workflows.

Demystify loop engineering by expanding a react agent into a five-step loop with two models, self-prompting, and memory shared across agents to refine outputs.

Configure a Gemini three notebook with internal Google search, add an external web search tool and a weather API, and set up API keys and environment variables for LangChain v1.

Learn to use Google Gemini 3 and Gemini 2.5 models with Lang chain in Python, compare pricing and token usage, and implement system and user message flow to invoke Gemini.

Learn how AI message structure uses content, text, and content blocks to represent model outputs. Understand tracing, tokens, and how system, human, and tool messages shape responses.

Learn to stream a model's output with LangChain ChatModel, printing chunks in real time to reduce perceived latency, using Gemini two as the underlying model.

Upload files to the Gemini model context, create a 30-minute limb cache from uploaded file parts, and configure a Gemini cache with a financial-analyst system instruction and display name.

Explore production context caching to optimize costs and speed, using cached conversations, unique context identifiers, and Gemini-based models to compare Q1 and Q2 2024 revenue and generate infographics.

Learn how to enable short-term memory by storing chat histories in sqlite, including setting up a sqlite3 connection, a multi-threaded financial agent db, and attaching a checkpoint to persist conversations.

Use guardrails and PII detection with LangChain to redact emails, URLs, IPs, and API keys before the model or after the tool call, with reusable PII middleware for end-to-end privacy.

Generate structured JSON outputs using a Pydantic base model with keys like company, stock symbol, current price, analysis, and recommendation. Use this schema to guide agent responses for fixed-key data.

Learn to craft a structured system prompt for a finance research assistant, connect Yahoo Finance MCP tools, and use stock data like get stock info and historical prices.

Connect Yahoo Finance MCP tools to a Gemini-based agent, building async finance research workflows that invoke tools, assemble messages, and return Apple stock data and latest news.

Create a stock researcher agent using a structured system prompt. Leverage available tools like the finance researcher tool and to-do list middleware to analyze stock performance and provide data-driven recommendations.

Master data extraction from pdfs using the doc link library to convert images, text, and tables into markdown, with metadata filtering and vector db retrieval.

Plan and implement an automated pdf data extraction pipeline with doc ling, converting pdfs to markdown, images, and tables, and organizing outputs into markdown, images, and tables directories.

Design a production-ready document processing pipeline that stores raw data, markdown, images, and tables by pdf name with page-wise context and enforces images of at least 500x500.

Extract metadata from file names to populate document type, company name, fiscal year and quarter, plus runtime metadata like page number and content type for deduplicated vector db storage.

Identify image items in a document with the doc converter, extract full-page images, and save them alongside page numbers. Process single PDFs first, then scale to multiple files.

Extract tables from markdown by identifying lines starting with a pipe operator, capture two preceding paragraphs as context, and save tables with page numbers for LM inference and data generation.

Verify data integrity by confirming PDF extraction into markdown, tables, and images across multiple companies, ensuring accurate ingestion and preparing image descriptions with a vision model for text data.

Load an image as base64 to generate its description. Read, buffer, and encode the image as png base64, then embed it in a prompt for the LLM.

Learn to invoke a language model with structured messages and system prompts to generate image descriptions using Gemini, including a reusable method that processes image paths and outputs markdown data.

Generate image description and save it as a markdown file in an output directory, using the image path metadata (company and document name), creating folders as needed and avoiding duplicates.

Generate and save descriptions for all images in the images directory, converting them to markdown. Track progress and ingest multi-modal data into a vector space for Rag retrieval.