CORTEX
All projects

Fuze

ONGOING

A bookmark manager that recommends saved pages for the project you are building, with its retrieval quality and latency measured.

Personal project, ongoing · Jul 2025 – present · 681 commits

THE PROBLEM

Saved bookmarks and posts pile up across tabs and apps and can't be searched by meaning, and a project needs the few saved pages that help with it, not a list of everything loosely related.

HOW IT WORKS

  • You create a project and tasks; the engine builds an intent from your own words and ranks saved pages by their best matching passage. Pages are split into overlapping chunks, embedded with bge-small (384 dimensions) and stored in PostgreSQL with pgvector, searched with HNSW indexes and diversified with MMR.
  • RQ workers on Redis fetch and embed pages in the background; Gemini adds a content profile per page and refines the intent, in jobs that no request waits for. A small cross-encoder flags weak matches, which the page folds behind a toggle.
  • A React frontend on Vercel, a Flask API in a Docker container on a free Hugging Face Space (gunicorn with gevent, workers under supervisord), Supabase Postgres, Upstash Redis and a Manifest V3 browser extension.

THE PIPELINE

  1. Save a pageFrom the browser extension, which can send the text the browser already rendered, or from a URL.
  2. Acquire the textOrdered tiers: extension capture, official site APIs, the backend's own fetch, a remote reader, then the Wayback Machine, with SSRF and robots.txt checks first.
  3. RQ workers on RedisSplit the text into overlapping passages of about 200 words and embed them with bge-small (384 dimensions); Gemini adds a content profile.
  4. PostgreSQL with pgvectorPassages and profiles stored with HNSW indexes; every vector records its embedding model.
  5. Retrieve and rankNearest passages for your project and task, diversified with MMR; a cross-encoder flags weak matches.
One engine ranks by meaning only. The page text is acquired without a browser on the server, and a weak match is flagged, never silently dropped.

ENGINEERING EVIDENCE

  • On BEIR SciFact (300 queries, run through the production search code) passage retrieval scores nDCG@10 0.718. On a long-document variant it scores 0.504 against 0.016 for one embedding per document. ADR-007-intent-engine.md ↗
  • Live semantic-search latency on the deployed Space, 30 samples per endpoint: p50 3,026 ms down to about 1,700 ms (1,661 at the best run, -45%) and SQL statements per search from about 10 to 1, with each run recorded and the unexplained remainder written down. README.md ↗
  • A cross-encoder rerank was measured and left off for search: it moved nDCG@10 by -0.008 and +0.004 for about 700 ms a query. It is used only to flag weak matches. ADR-007-intent-engine.md ↗
  • The free Redis plan's 500,000 monthly commands ran out. Removing the scheduler and per-bookmark polling brought an idle 10-minute sample to 2 commands. OPERATIONS.md ↗
  • The passage table has an HNSW index (m=16, ef_construction=64) created in an Alembic migration. 0011_content_chunks.py ↗
  • Backend CI, Frontend CI, a Docker build check and the Space sync all pass on the latest commit (5e47109).

DECISIONS & INVESTIGATIONS

WHAT ISN'T DONE

  • Throughput under load is not measured on the live Space: its benchmark is 30 samples per endpoint from one machine in India. Recommendation quality is measured offline, not on real users' feedback.
  • About 900 ms of each semantic search is unexplained. Postgres runs the statement in about 2 ms, yet a trivial query from the Space to the database (Supabase, Singapore) costs 213-228 ms. Moving the database closer is written up and deferred; the Space's own region is not documented.
  • SciFact is scientific abstracts, not bookmarks, so no benchmark on real saved pages exists. The weak-match floor was calibrated on one 238-bookmark library over 15 requests, 23% of useful results were also flagged, and the repository does not record who assigned the relevance labels.
  • The browser extension has not been exercised end to end in real Chrome against the deployed backend.
  • The earlier two-stage recommendation pipeline, its shadow mode and the golden-set regression gate were deleted on 12 Oct 2026, and the decision records about them carry a note saying so.
  • Built with an AI coding assistant: 11 of the 681 commits carry a Claude co-author trailer, all from before 12 Oct 2026, so that count understates it.

NEXT STEPS

Each one comes from a gap listed above. It says what fixing the gap would take; it is not a promise.

  • Move the database closer to the Space, then re-run the live benchmark to see how much of the unexplained 900 ms was distance.
  • Record who assigned the weak-match relevance labels, and re-check them with independent raters before quoting the figures as a benchmark.
  • Add a retrieval benchmark on real saved pages, since SciFact only stands in for them.

STACK

  • Python
  • Flask
  • PostgreSQL
  • pgvector
  • Redis
  • RQ
  • SentenceTransformers
  • Gemini
  • React
  • Vite
  • Alembic
  • GitHub Actions