Summariser

Paste a URL. Firecrawl turns it into markdown, a crewAI crew summarises it, and you can chat with the source.

• Scrape: Firecrawl v2 /scrape (markdown, main content only)
• Crew: Editorial Summariser (scaledown/summarize) gets the full scraped markdown, as-is, and writes an overview plus key points sized to the article. Falls back to a direct ScaleDown call if the crew fails.
• Chat: google/gemma-4-31b-it:free over the summary plus source text, with short-term history per session. If it is rate-limited or returns nothing, the request automatically retries the next free model in CHAT_FALLBACK_MODELS and reports which one answered.
• Groundedness badge: after each answer, baseten/mammoth-1:free returns a probability that the source supports the answer via OpenRouter's Decision API.

Why mammoth doesn't write the chat

baseten/mammoth-1:free is a decision model, not a text model. It accepts a state plus typed questions (noul, choice, score) at POST https://openrouter.ai/api/alpha/decisions and returns calibrated probabilities with no generated prose. Asking it to answer a chat message would fail, so it is used for what it is for: scoring whether each answer is grounded in the source. Set MAMMOTH_CONFIDENCE=0 to skip that call.

Setup

Requires Python 3.10–3.13 and uv.

cp .env.example .env      # add your keys
uv sync --python 3.13
uv run uvicorn app.main:app --reload

Open http://127.0.0.1:8000.

You need:

• OPENROUTER_API_KEY — https://openrouter.ai/keys
• FIRECRAWL_API_KEY — https://firecrawl.dev/app/api-keys

scaledown/summarize bills at $0.04/M input tokens; the Gemma chat model and mammoth-1 are free tiers.

Configuration

Variable	Default	Notes
SUMMARISER_MODEL	scaledown/summarize	crewAI summariser agent
CHAT_MODEL	google/gemma-4-31b-it:free	primary chat model
CHAT_FALLBACK_MODELS	poolside/laguna-s-2.1:free,dots-studio/dots-3-note-preview:free,recursive/flint-v1.0:free,cohere/north-mini-code:free,openrouter/free	comma-separated fallbacks
MAMMOTH_MODEL	baseten/mammoth-1:free	groundedness decisions
MAMMOTH_CONFIDENCE	1	0 disables the badge
MAX_SOURCE_CHARS	200000	scrape truncation
CHAT_CONTEXT_CHARS	120000	source text sent to chat
SESSION_TTL_MINUTES	180	in-memory session lifetime

API

• POST /api/ingest — {"url": "https://..."} → {session_id, title, url, summary, markdown, mode}
• POST /api/chat — {"session_id": "...", "message": "..."} → {answer, grounded, model}
• GET /api/session/{id} — reload a session
• GET /api/health — key and model status

Deploy

Pushing to main runs the tests and then deploys with GitHub Actions: rsync to /srv/summariser, docker compose up -d --build, health check on http://localhost:8011/api/health. Required repository secrets:

• DEPLOY_HOST — public IP of the Hetzner box
• DEPLOY_USER — yushan
• DEPLOY_SSH_KEY — private deploy key
• DEPLOY_KNOWN_HOSTS — ssh-keyscan output for the host

The server keeps its own .env (excluded from sync). The container binds to loopback and is exposed only inside the tailnet by tailscale serve on HTTPS port 8443.

Notes

Sessions are in-memory, so a restart or a second uvicorn worker loses them. For a real deployment, put the session store in Redis/SQLite and add auth and rate limiting on /api/ingest (each call spends Firecrawl and ScaleDown credits).

Tests

uv run pytest