digest-bot

I built a RAG chatbot that answers questions over a rolling window of RSS/changelog feeds instead of a static, one-time-indexed corpus. Most RAG tutorials index a fixed set of documents once and stop there — I wanted to work through the parts they skip: keeping the index fresh as new articles arrive, deduping across re-polls, and weighting retrieval toward what's recent instead of just what's semantically closest.

How it works

1. Ingest (ingest/) — poll RSS feeds, extract readable article text, dedup by GUID/URL.
2. Index (index/) — chunk articles, embed chunks, store in a local vector DB with published_at + source metadata.
3. Retrieve (retrieve/) — I blend vector similarity with a recency-decay weight so a fresher, slightly-less-similar chunk can outrank an older, closer one. This is recency-weighted vector search, not hybrid (keyword + vector) retrieval — there's no BM25/keyword component.
4. Generate (generate/) — answer questions, forcing citations back to source articles; explicitly say "I don't have recent info" when retrieval is empty.
5. Eval (eval/) — a harness that measures retrieval recall@k and citation correctness against a hand-written question set, so the answer to "does this work" is a number rather than vibes. The question set itself ships as a template, because an article id only means something against the feeds you actually poll — see Eval.

Quickstart

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env   # add your ANTHROPIC_API_KEY

python -m ingest.poll       # pull + store new articles
python -m index.build       # chunk + embed + index
python -m generate.chat     # ask questions in a CLI loop

Only the API key lives in .env; the feeds are in config/feeds.yaml. The first ingest.poll takes a minute or two — it fetches the full text of every article it takes, and prints a per-feed count as it goes. index.build downloads the embedding model on its first run.

Config

Feeds live in config/feeds.yaml. Add/remove RSS URLs there, along with max_articles_per_feed — some feeds publish their whole history in one document (the OpenAI blog is over a thousand entries), and since every entry costs a page fetch, an unbounded poll is a fifteen-minute one. The default takes the 25 newest per feed; set it to 0 for no limit.

Eval

python -m eval.run

Runs the question set in eval/questions.yaml against the current index and reports retrieval recall@k and whether generated answers cited the right source.

eval/questions.yaml ships as a template rather than as my own question set, because an expected_article_id is only meaningful against the feeds it came from — point the bot at a different beat and every id in it is dead. Fill it in from data/articles.jsonl after the first index build. Until then eval.run says so and exits, instead of scoring the placeholder and reporting a 0% that looks like a retrieval failure.