DocChat

DocChat is a small, production-minded local RAG chatbot: ask questions over your own PDFs, fully offline, with Ollama as the LLM and ChromaDB + sentence-transformers for retrieval.

Prerequisites

• Python 3.11+ (required — DocChat is not tested on older interpreters)
• Ollama installed and running locally
• The mistral model pulled: ollama pull mistral

Before creating the venv, confirm the interpreter: python --version should report 3.11.x or newer. On Windows, if python still points at an older release, use the Python 3.11+ launcher instead, for example py -3.11 -m venv .venv, so pip install pulls pre-built wheels and you avoid a from-source greenlet build that requires Microsoft C++ Build Tools.

Setup

1. Open a terminal in the docchat directory (the folder that contains backend/, frontend/, and requirements.txt).
2. Create a virtual environment (recommended) and install dependencies:

 python -m venv .venv
 .venv\Scripts\activate
 pip install -r requirements.txt

On Windows, if python is not 3.11+, prefer: py -3.11 -m venv .venv then activate and pip install as above. 3. (Optional) Copy .env.example to .env if you want to document local overrides for your environment. 4. Start the API (from the docchat directory):

 python -m uvicorn backend.main:app --reload --host 127.0.0.1 --port 8000

5. Open the UI in a browser: http://127.0.0.1:8000/

Usage

1. Copy PDFs into data/pdfs/ (subfolders are fine; ingestion is recursive).
2. Click Ingest PDFs in the sidebar (or POST /ingest).
3. Choose Langgraph or Langchain as the pipeline, then chat.

The vector database is written to chroma_db/ on disk and reused across restarts.

API

Method	Path	Description
GET	/health	Liveness probe: { "status": "ok" }.
GET	/status	Ollama reachability (/api/tags), indexed chunk count, active LLM and embedding model names.
POST	/ingest	Body: { "folder_path": "<optional path>" }. Defaults to configured data/pdfs. Loads PDFs, chunks, embeds locally, persists to Chroma. Returns status, files_processed, chunks_created.
POST	/query	Body: { "question": "<text>", "use_graph": true }. If use_graph is true, runs the LangGraph StateGraph pipeline; if false, runs the LangChain retrieval + stuff-documents chain. Returns answer, sources, chunks_used, and pipeline (langgraph or langchain).

Errors return JSON with a detail field (FastAPI default). If nothing has been ingested yet, /query responds with 400 and: No documents found. Please ingest PDFs first via POST /ingest.

Architecture

DocChat ships two interchangeable RAG paths over the same Chroma vector store:

1. LangGraph (use_graph: true) — A StateGraph with two nodes: retrieve (similarity_search with TOP_K) and generate (prompt + ChatOllama). This demonstrates explicit graph orchestration, easy extension (e.g. re-ranking, tool calls, branching), and clear state (question, retrieved_docs, answer, sources).
2. LangChain chain (use_graph: false) — create_retrieval_chain + create_stuff_documents_chain with the same retrieval width and system instructions. This is the classic concise RAG stack for comparison and interoperability.

Both use the same embeddings, chunking policy, and Ollama model configured in backend/config.py.

Project structure

docchat/
├── backend/
│   ├── __init__.py
│   ├── main.py        # FastAPI app, CORS, routes, lifespan
│   ├── ingest.py      # PDF load → split → embed → Chroma
│   ├── rag_chain.py   # LangChain retrieval + stuff chain
│   ├── graph.py       # LangGraph StateGraph RAG
│   └── config.py      # Paths, models, chunking, top-k
├── frontend/
│   └── index.html     # Single-file dark UI
├── data/
│   └── pdfs/          # Drop PDFs here
├── chroma_db/         # Created by Chroma after first ingest
├── requirements.txt
├── .env.example
└── README.md