mirror of
https://github.com/NeuralNomadsAI/CodeNomad.git
synced 2026-10-29 10:02:49 +00:00
8e8ea0653e
## Summary
Follow-up to #1161 and #1164. In a large project, the ranked project
search covered only the **oldest** messages: the scan walked rowids
upward under a 1.5 s budget, so recent conversations were never reached,
and identical searches covered 5,000 to 7,000 messages depending on
load. Each 1,000-row statement also sorted and evaluated its whole chunk
before returning its first row, blocking the daemon for about 70 ms at a
time.
- **Newest first:**
- The scan streams 128-row ranges from an O(1) horizon, without a
temporary sort.
- It covers a deterministic window of the 16,384 most recent messages.
- **Cheaper per row:**
- A necessary UTF-8 byte check on the raw row rejects messages that
cannot contain a query term before any JSON is decoded.
- Only possible matches pay for the exact readable-text mask.
- The existing time, candidate, session and judge guards remain.
- **40 results** in the project search window; the agent tool keeps 10
by default and 20 at most.
- **Search further back:**
- When older history remains, it ranks the next older window for the
same query and options through an authenticated per-location
continuation.
- The continuation is bound to scope, query, options and the original
horizon, so later appends never enter it.
- Older windows replace the list, because ranking scores are
window-local.
- The coverage note names the examined range in all ten locales.
- Query, scope or option changes, closing and leaving the session reset
it.
- The conversation scope is unchanged.
## Measurements (synthetic 40,000-message corpus, `node
scripts/benchmark-ranked-history.mjs
|
||
|---|---|---|
| .. | ||
| features/wake-lock | ||
| guides | ||
| help | ||
| screenshots | ||