Skip to content

BACKLOG.md

Latest commit

 

History

History
63 lines (60 loc) · 5.86 KB

BACKLOG.md

File metadata and controls

63 lines (60 loc) · 5.86 KB
Raw

Backlog

Open work only. A finished item is deleted; its record is the merged pull request.

  • Serve relay, what v1 leaves out (the serve relay): a Windows named pipe with an owner-only access list (NEED-1405; off on Windows now); forwarding an ask question to the client's terminal instead of requiring --yes; waiting for a serve that is still starting instead of failing on the messenger's own lock; per-watcher buffer budgets for a relayed watch (max-cli's server/subscribers.ts); mcp --http through serve; a path type on options instead of the placeholder naming rule (DEBT-124).

  • One message search without a language switch — Lucene syntax with legacy's forgiving matching, matchers run in parallel, merged and ranked against the query, each hit labelled with how it matched (the legacy search and --language are already gone). Ranking test set and an owner-approved plan come first: combined-search-handoff.md. Approved plan for the narrow first step (B): typo-search-default-plan.md.

  • Messenger feature gaps — albums, archiving, typing and location, MAX cloud password and video notes. Proposed, not claimed: messenger-feature-gaps.md.

  • Chat briefs and news digests — decisions, requests and open questions per chat for a time window, and a cited digest of selected news feeds, prepared for the owner's agent. Not designed beyond a proposal; needs the owner's scope before a plan.

  • Mail embedding: what is not measured yet — a model-based line classifier for cleaning, and a run of bench/mail-quality on the owner's own mailbox, which would stay in cli-private.

  • A reranker for note and document search, opt-in — memo and drive only fuse a word list and a meaning list, and miss paraphrases (2 of 5 in the top 3 against qmd's 4 of 5: local-documents). A query-passage model as a separate download, like embeddinggemma, run only in a long-lived process (MCP server); never on by default. The message-search rerankers did not ship — the rule one changed nothing in the dev sweep, e5 cosine lowered recall, the cross-encoder missed held-out gates at 3–4 s (combined-search.md). Measured on the local-documents queries: mmarco (130 MB) lifts memo's notes from 0.72 to 0.84 first place and every paraphrase into the top 3, ~2 s a query warm; on Drive documents the gain is within noise; qmd's Qwen3 is no better and three times slower (rerankers). One run of 25 queries per corpus, no held-out set — the message cross-encoder also looked good on dev and failed held-out. Before building: a fresh query set, partly the owner's, run once with the setup frozen. Owner (NEED-1004): behind its own flag, off by default; score alone or qmd's blend is decided after the second, frozen run.

  • No query-expansion model — qmd rewrites each query with a 1.7B model. Our callers are agents that can search again in other words; the skills say so. A model for this stays optional and off by default, if ever.

  • EmbeddingGemma for documents — ours already ships it as an option; on documents it was not compared with e5-small. Measured together with the reranker; the default for notes changes only on that result.

  • Generated tags for notes, documents and messages, searchable at once — claims kept apart from the owner's tags, below them in tag:, kept as a status on taggings; tags in any language with a transliterated form for the duplicate check; a model or an agent may set them. Owner: do it all, vectors too (NEED-1005 A). Proposal: generated-tags-plan.md.

  • One registry for what gets a vector — short fields (tag names, titles) and long texts each embedded one way; only registry fields get vectors; a test fails on writes around it. Owner direction 2026-10-13. Proposal: embedding-registry-plan.md.

  • Mentions: phones, e-mail, links grouped by site, documents, names and places found in text — one entity with a kind, reviewed by an agent and the owner, linked to people, organizations, documents and places; rules, no model, by default. Owner direction NEED-1006 A. Proposal: mentions-plan.md.

  • Embedding throughput — notes, mail and documents embed in one session on min(8, cores) threads (src/services/notes-search.ts), 8 of 24 cores here; one text per session.run (src/embeddings/embed.ts). The worker benchmark split the same 8 threads between workers (1.04–1.1×); more sessions with 8 threads each were not measured (two Gemma runs side by side here: roughly 1.3× together, one batch, not a measurement). On 1,200-character document passages e5-small embedded 11–15 a second and EmbeddingGemma ~1.6, not the 31 and 2 in the model table. Measure, in order: more sessions over all cores (~700 MB each), batched inputs, a native ONNX Runtime (larger install).

  • Search start-up — memo and drive take 1.2–2.4 s a search from the CLI; the same vector search in a warm process takes 7 ms, so most of the CLI time is probably start-up and model loading. Measure where the CLI time goes (process start, store open, model load) before tuning search.

  • Speech throughput — voice messages are heard one after another, the recognizer is opened and freed for each (src/speech/transcribe.ts hearLocally), on one thread (src/speech/recognize.ts). Measure one recognizer per pass, then several workers over files (each holds its own model).