FastAPI + RAG demo
A retrieval-augmented Q&A agent over the fastapi/full-stack-fastapi-template docs.
Reproduce
cd /tmp
git clone https://github.com/fastapi/full-stack-fastapi-template.git
cd full-stack-fastapi-template
git checkout <pinned-SHA> # see examples/hero/fastapi_rag/run.sh
uvx harnessforge init --no-llm --blueprint rag-agent
What gets written
The repo gains:
AGENTS.mddescribing the RAG loop, conventions, and definition of doneSOUL.mdsetting the "careful research assistant" voiceTOOLS.mdrecommending qdrant + filesystem + fetchMEMORY.mddefining the three-layer memorySKILLS/chunk-and-embed/SKILL.mdfor ingesting the FastAPI docsSKILLS/retrieve-and-rerank/SKILL.mdfor top-k + cross-encoderSKILLS/answer-with-citations/SKILL.mdfor citation-enforced answersSKILLS/eval-recall-precision/SKILL.mdfor the eval set runner- Adapter files for Claude Code, Cursor, Continue, Windsurf, Codex
Verify
Should exit 0 with the structure check passing (citations + schema skip until an agent actually produces output).
Now use it
Open the repo in Claude Code. Claude reads .claude/CLAUDE.md, learns the project type, the loop, the forbidden paths, and the citation contract — without any per-session prompting.
Or run Hermes in the same dir — Hermes reads SOUL.md and SKILLS/ and starts with the same ground truth.
Why this demo
- Multi-file output over a recognizable stack (FastAPI + React + Postgres)
- Demonstrates cross-agent interop: Claude Code, Cursor, OpenHarness all read the same
SKILLS/andAGENTS.md - Real corpus to ingest (the repo's own docs)