07.01 — What Are We Building?

Overview

This lesson introduces the Documentation Helper project — a full-stack RAG application that lets users ask natural language questions about a package's documentation and receive accurate, grounded answers with source citations. We're building this end-to-end: the ingestion pipeline, the retrieval chain, and a chat-based frontend.

---

The Project: Documentation Helper

flowchart LR
    USER["👤 User:\n'How do I use ChatOpenAI\nwith streaming?'"]
    APP["🖥️ Documentation Helper"]
    ANSWER["✅ Grounded Answer:\n'To use ChatOpenAI with streaming,\nuse the .stream() method...'\n\n📎 Sources:\n- docs.langchain.com/chat/streaming\n- docs.langchain.com/models/openai"]

    USER --> APP --> ANSWER

The application takes any package's documentation (in this case, the LangChain documentation), ingests it into a vector store, and enables question-answering over it — with source citations so the user knows exactly where the answer came from.

---

The Four Parts

The project is built in four phases:

flowchart TD
    P1["1️⃣ Ingestion\nDownload docs → chunk → embed → store"]
    P2["2️⃣ Retrieval Chain\nQuery → search → augment → generate"]
    P3["3️⃣ Frontend\nStreamlit chat interface"]
    P4["4️⃣ Memory\nChat history for multi-turn conversations"]

    P1 --> P2 --> P3 --> P4

    style P1 fill:#4a9eff,color:#fff
    style P2 fill:#10b981,color:#fff
    style P3 fill:#8b5cf6,color:#fff
    style P4 fill:#f59e0b,color:#fff

Phase	What We Build	Key Technologies
Ingestion	Download the LangChain docs, chunk every page, embed with OpenAI, store in Pinecone	Tavily Crawl, RecursiveCharacterTextSplitter, OpenAIEmbeddings, PineconeVectorStore
Retrieval	Create a LangGraph agent with a retrieval tool that searches the vector store	create_agent, @tool, as_retriever, initChatModel
Frontend	Build a chat interface where users type questions and see answers with sources	Streamlit, session state, chat messages, source expanders
Memory	Add conversation memory so the chat can reference previous questions	Streamlit session state

---

What Makes This Different from Section 06

Aspect	Section 06 (RAG Essentials)	Section 07 (Documentation Helper)
Data source	Single text file (one blog post)	Entire website (hundreds of pages)
Crawling	Manual file copy	Automated web crawling (Tavily)
Scale	~20 chunks	~6,500 chunks
Retrieval	LCEL chain (deterministic)	LangGraph agent with tool (agentic)
Frontend	None (CLI only)	Streamlit chat UI with sources
Rate limiting	Not an issue	Critical — concurrent batch processing

---

Topics Covered

This section dives deep into production-relevant topics:

• Web crawling — automated documentation scraping with Tavily
• Concurrent processing — async batch extraction and indexing
• Rate limiting — handling API throttling with retry strategies
• Text splitting — RecursiveCharacterTextSplitter for semantic chunking
• Agent-based retrieval — LangGraph agent with a custom retrieval tool
• Content vs artifact — separating LLM context from application data
• Streamlit — Python-based rapid UI prototyping
• Production patterns — analysis of Chat LangChain (LangChain's own production RAG app)
• RAG architectures — two-step, agentic, and hybrid approaches

---

Summary

We're building a Documentation Helper that: 1. Crawls the LangChain documentation website automatically 2. Ingests it into a searchable vector store (Pinecone) 3. Answers questions using a LangGraph agent with a retrieval tool 4. Cites sources so users can verify the answer 5. Runs in a Streamlit chat interface for easy testing and QA