B2A platform: provenance-grade Portugal residential property data + signed JWS verification for AI agents.

Authoritative docs: attesta-architecture.md · attesta-agent-prompt.md

Quick start

bash

make install
cp .env.example .env
make db-up                 # Postgres + pgvector
make migrate               # apply Alembic migrations
make dev                   # API at :8000

Endpoint	URL
Health	`/healthz`
REST API	`/docs`
MCP	`/mcp`
JWKS	`/.well-known/jwks.json`
Agent discovery	`/llms.txt`

Pipeline (local)

bash

make ingest    # attesta-ingest --fixture (bundled JSON corpus)
make resolve   # attesta-resolve
make test

Ingestion

Default: bundled fixtures in tests/fixtures/ — no external scraper credentials needed.

bash

make ingest                              # same as attesta-ingest --fixture
attesta-ingest --fixture                 # explicit
attesta-ingest --source idealista_pt     # optional: Apify actor (needs APIFY_* in .env)
attesta-ingest --source imovirtual       # optional: Apify actor

Live portal collectors implement app/ingestion/base.py (BaseCollector). Apify-backed Idealista/Imovirtual collectors ship as optional plugins; add a new source by subclassing, not refactoring the pipeline. See docs/decisions/008-fixture-first-ingestion.md.

To grow production corpus: pick a collector strategy (build Apify actors, direct HTTP, licensed feed, CSV import), wire env, and schedule the ingest worker on Railway.

CI

GitHub Actions runs ruff, mypy (advisory), pytest on SQLite + Postgres.

Decisions

Architecture Decision Records: docs/decisions/
Cursor rules: .cursor/rules/

Milestone status

M0 — Skeleton, CI, docker-compose, alembic scaffold
M1 — Schema, ingestion worker, fixtures, PII tests
M2 — Resolution matcher, read API
M3 — MCP tools, metering choke point, guardrails
M4 — Verification pipeline, JWS/JWKS, evidence API, eval harness (62 claims), attestation spec
M5 — Railway 3-service config, demo + verify scripts, OpenAPI export

Verification

bash

make eval                    # 62 labeled claims, 75% threshold
python scripts/verify_jws.py <jws> --local
python scripts/demo.py       # E2E against local API
make openapi                 # writes public/openapi.json

LLM (optional, provider-agnostic via LiteLLM)

Set LLM_ENABLED=true and configure models + prompt file paths in .env (see .env.example).
Uses LiteLLM — point LLM_PARSE_MODEL / LLM_ADJUDICATE_MODEL at any supported provider (anthropic/, openai/, gemini/, azure/, groq/, etc.).
Provider API keys follow LiteLLM env conventions — set only the key(s) for your chosen provider(s).
See distribution/llm-providers.md for examples and cost estimates.