lore-backend

The FastAPI service behind Lore — an engineering team's decision memory. Captures the why behind merged pull requests (via a GitHub App) and answers /why questions with cited, sourced answers.

A LangGraph agent that can go fetch what it doesn't have, over three memory tiers, behind a guardrail that won't ship an uncited answer. Runs entirely on free and self-hosted services.

Companion repos: lore-cli (the npx lore git-hook CLI that captures commit Why: trailers) and lore-vscode-extension.

For the design and the reasoning behind it, see ARCHITECTURE.md and ADR-0001.

Quickstart (self-host)

cp .env.example .env   # works as-is in MOCK mode — no keys required
docker compose up
curl localhost:8000/health

Postgres, Qdrant, the API and the worker, in one command.

• MOCK mode (no GROQ_API_KEY): answers come from token-overlap search over a curated seed corpus. Deterministic, no external calls — good for demos and CI.
• LIVE mode: add GROQ_API_KEY and restart for the agent loop, real retrieval, and real GitHub ingestion.

DATABASE_URL is required in both modes — the control plane and episodic memory live in Postgres regardless.

Install from PyPI

Published at pypi.org/project/lore-backend.

docker compose up stays the recommended way to run Lore — it brings its own Postgres and Qdrant. The package is for embedding the service in an existing deployment, or importing the memory layer directly:

pip install lore-backend
export DATABASE_URL=postgresql://lore:lore@localhost:5432/lore
lore-backend                 # the API server (--host/--port/--reload)
lore-backend-worker          # the background job worker

Both scripts read the same environment as the Docker stack (see .env.example); you supply Postgres and Qdrant yourself. Prompts and migrations ship inside the package, so procedural memory and the migration runner work from an install with no checkout.

Tech stack

Layer	Choice	Cost
Web framework	FastAPI + Uvicorn	—
Orchestration	LangGraph StateGraph	OSS, MIT
Agent LLM	Groq — Llama 3.3 70B	free tier
Summarizer LLM	Groq — Llama 3.1 8B	free tier
Embeddings + reranking	fastembed (ONNX, CPU)	$0, no API
Semantic memory	Qdrant (or pgvector)	free / self-host
Episodic memory + control plane	PostgreSQL	self-hosted
Job queue	Postgres FOR UPDATE SKIP LOCKED	no broker
Tracing + eval	Langfuse	OSS, self-host
Auth	DB-backed API keys, sha256-hashed	—
Testing	pytest, ruff	—

No new cloud bill. Nothing here requires an account anywhere.

How an answer gets made

question → agent ─┬─(needs more)→ tools → agent
                  └─(has enough)→ guardrail → answer + citations

The model decides whether the Canon covered the question or whether it needs to go read the PR itself. Capped at 4 hops. Before anything reaches the user, the guardrail checks every citation against what retrieval actually returned — an answer citing a PR that was never retrieved does not ship.

Memory is three tiers, because they answer different questions: procedural (how to behave — prompts/*.md, git-versioned), semantic (durable distilled decisions — vector search), and episodic (dated events and past answers — SQL, ordered by time).

Details in ARCHITECTURE.md.

API

Method & path	Purpose
POST /v1/why	Core Q&A — returns answer, sources, path, hops, guardrail, trace_id
GET /v1/why/history	Recently answered questions for this Canon
GET /v1/canon, GET /v1/memories	Cursor-paginated dump of the Canon
POST /v1/lore	Free-text search, no composed answer
POST /v1/ingest/seed	Load the seed corpus (LIVE mode)
POST /v1/inscribe	CLI writes a commit's Why: (idempotent)
GET /v1/backfill/status, POST /v1/backfill/run	Installation backfill
POST /v1/keys, DELETE /v1/keys/{id}	API keys (admin-secret-gated)
GET /health	Mode, active path, loaded prompts, tracing status
GET /metrics	Prometheus-format counters
POST /webhook/github	GitHub App webhook receiver

Eval

python -m lore_backend.eval.harness              # score the active path
python -m lore_backend.eval.harness --compare    # v1 pipeline vs v2 agent, same store
python -m lore_backend.eval.harness --judge      # add LLM-as-judge (LIVE only)

A golden question set scored for citation accuracy and relevance. Deterministic in MOCK mode, so tests/test_eval_harness.py holds a hit-rate floor as part of pytest, and the harness itself exits non-zero when the gate fails — run it before merging anything that touches retrieval. The LLM judge is observe-only until its scores have been checked against human reading.

The --compare mode is why the v1 pipeline is still in the tree: the loop has to out-perform something. Note that the agent path is non-deterministic — one run is a sample, not a measurement.

Tracing (optional)

docker compose -f docker-compose.yml -f docker-compose.langfuse.yml up

Open http://localhost:3000, create a project, put its keys in .env as LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY, restart the backend. Until then every trace call is a no-op — nothing in the request path depends on Langfuse being reachable.

Configuration

Everything lives in .env.example with safe defaults. The knobs worth knowing:

Variable	Default	What it does
AGENT_LOOP_ENABLED	true	false falls back to the v1 pipeline
AGENT_MAX_HOPS	4	Tool-call budget per question
RERANK_ENABLED	true	Cross-encoder rescoring of the shortlist
VECTOR_STORE	qdrant	Or pgvector to reuse the same Postgres
CONSOLIDATE_AFTER_N_EVENTS	25	When the summarizer distils a tenant's backlog
JUDGE_ENABLED	false	LLM-as-judge in the eval report
EVAL_HIT_RATE_FLOOR	0.9	What the eval gate enforces

Local dev (no Docker)

python -m venv .venv && . .venv/Scripts/activate  # or source .venv/bin/activate
pip install -r requirements.txt
# Needs a reachable Postgres — `docker compose up postgres` or your own.
uvicorn lore_backend.main:app --reload --port 8000
# in another terminal:
python -m lore_backend.jobs.worker

Testing

pytest
ruff check .

Tests need a live Postgres (DATABASE_URL) — they run real migrations and truncate between cases rather than mocking the database. The agent, judge and summarizer are driven by scripted fakes, so the suite makes no network calls and needs no API keys.

tests/test_kafka_ingestion_demo.py needs the demo's own optional dependency (lore_backend/examples/kafka_ingestion/requirements.txt).

Deploying the GitHub App

Point the App's webhook URL at <your-host>/webhook/github and set GITHUB_WEBHOOK_SECRET / GITHUB_APP_ID / GITHUB_APP_PRIVATE_KEY. The payload shapes handled are in lore_backend/ingestion/webhook_handler.py.