Chat with your knowledge base

A conversational RAG agent — LangGraph + OpenRouter, built on Fast Dash's native chat mode.

▶ Live demo

Share a web page, PDF, YouTube video, or pasted text by chatting, then ask questions — the agent ingests your sources, retrieves the relevant passages, and answers with inline citations. You watch it work: the "adding source" and "searching" steps show up as cards in the transcript.

A modern rebuild. The original version was a form (paste your OpenAI key + URLs + a query) powered by Embedchain, which re-built the whole vector store on every query. This rewrite drops Embedchain and OpenAI entirely: it's a LangGraph ReAct agent on OpenRouter streamed through Fast Dash chat, with a per-session knowledge base and local, key-free embeddings.

How it works

The entire app is one line — a compiled LangGraph graph handed to Fast Dash's chat mode:

app = FastDash(callback_fn=build_graph(), chat=True, title="Chat with your knowledge base")

The agent has two tools:

• add_source — load a URL / YouTube link / PDF / text, chunk it, embed it locally (FastEmbed), and store it in this session's vector store.
• search_knowledge — semantic retrieval; the agent answers only from the retrieved passages and cites them [1], [2].

Fast Dash's langstage bridge streams the agent's tokens and tool calls into the chat, and gives it multi-turn memory. OpenRouter runs the LLM; retrieval embeddings are computed locally (no OpenAI, no embeddings API key).

Run locally

uv sync
export OPENROUTER_API_KEY=...        # https://openrouter.ai/keys
uv run python -m app                 # http://127.0.0.1:8080

Optional: KNOWLEDGE_CHAT_MODEL (default anthropic/claude-haiku-4.5), EMBED_MODEL (default BAAI/bge-small-en-v1.5).

Deploy

Served by gunicorn (gthread, single worker) so the chat history and per-session knowledge bases live in one process. The FastEmbed model is baked into the image at build time. See the Dockerfile. Set OPENROUTER_API_KEY as a secret on your host.

Architecture

File	Role
app.py	FastDash(callback_fn=graph, chat=True) — the whole app
rag/graph.py	The ReAct RAG agent on OpenRouter (system prompt + tools)
rag/knowledge.py	Per-session vector store, source loaders, chunking, retrieval
rag/tools.py	add_source / search_knowledge tools (session-scoped via config)

Stack

Fast Dash (chat mode) · LangGraph · LangChain (OpenRouter) · FastEmbed (local embeddings) · web / PDF / YouTube loaders.

License

MIT — see LICENSE.