Skip to content
RAG Support Assistant

RAG Support Assistant

A customer-support agent that retrieves product docs, routes every answer through LangGraph, and keeps replies grounded in known sources.
See the example Reproduce E20 GitHub

Proof points

Section titled “Proof points”

Published docs hub

Docs are published under /RAG_Support_Assistant. English and Russian entry pages are maintained explicitly.

Seeded support KB

The demo KB starts from three seed documents: warranty, returns, and E10-E30 errors. E20 is present in errors_e10_e30.md.

Fixture-backed evaluation

The fixture set has 12 cases across error codes, password reset, warranty, installation, billing, and general support. The three-document demo KB does not cover every fixture category.

Traceable integration surface

The API catalog is tied to FastAPI route decorators, and /api/ask returns an answer with documents and citations.

One question’s path through LangGraph

Section titled “One question’s path through LangGraph”

auto

retry

Question

classify_complexity

transform_query

retrieve

grade_docs

generate

verify_facts

evaluate

route_or_retry

Answer + sources

The full flowchart, with all 12 nodes and conditional transitions, is generated automatically from agent/graph.py on the LangGraph state machine page.

What /api/ask returns for “How do I fix E20?”

Section titled “What /api/ask returns for “How do I fix E20?””
{
  "answer": "Error E20 is a water-drainage problem. Likely causes: a clogged drain filter, a kinked drain hose, or a faulty drain pump [1].",
  "sources": [
    { "source": "errors_e10_e30.md", "page_content": "E20 — water drainage problem …" }
  ],
  "citations": [
    { "index": 1, "doc_id": "errors_e10_e30.md", "title": "errors_e10_e30.md", "excerpt": "drain valve / filter …" }
  ]
}

The shape is fixed by the Pydantic models AskResponse, SourceInfo, and Citation in api/routers/conversation.py. For the full node-by-node walkthrough, see What it does.

Start here

Section titled “Start here”
  • What it does walks one E20 question through the graph.
  • Reproduce E20 gives the deterministic trust path: seed, ingest, ask, verify.
  • Try locally covers the local setup path.
  • Evaluation covers what is measured, how, and where the limits are.
  • GitHub contains the source.