@yurtsever/capsa 0.1.0-alpha.7 → 0.1.0-alpha.8

package/README.md

@@ -55,7 +55,8 @@ Once published, `npx capsa …` replaces the `node packages/cli/dist/index.js
 | Kind | Where | Ranking signals |
 |---|---|---|
 | Instruction files | `CLAUDE.md`, `AGENTS.md`, `.cursorrules`, `.cursor/rules/*.mdc`, `SKILL.md` | kind |
-| Runbooks / wiki | `docs/**/*.md`, `runbooks/**/*.md` | recency |
+| Docs / wiki | `docs/**/*.md`, `wiki/**/*.md` | recency, kind |
+| Runbooks | `runbooks/**/*.md`, `docs/runbooks/**/*.md` | recency |
 | Decisions | `docs/adr/**/*.md`, `decisions/**/*.md` | recency, kind |
 | Tickets | `tickets/**/*.md`, `docs/tickets/**/*.md` | **status** from checkboxes or `status:` frontmatter, recency |
 | Plans | `plans/**/*.md` | status, recency |
@@ -64,7 +65,7 @@ Adding a source is one adapter package (`packages/adapter-*`) — the scanner do
 
 ## How retrieval works
 
-Vector search (bge-m3 via Ollama, [sqlite-vec](https://github.com/asg017/sqlite-vec)) and full-text search (FTS5) are fused with reciprocal rank fusion, then weighted by what the source can honestly claim about itself: how recent it is, whether it is still live (`in-progress` beats `done`), and what kind of knowledge it is. Results are filled greedily into the token budget, at most two chunks per file.
+Vector search (bge-m3 via Ollama, [sqlite-vec](https://github.com/asg017/sqlite-vec)) and full-text search (FTS5) are fused with reciprocal rank fusion, then weighted by what the source can honestly claim about itself: how recent it is, whether it is still live (`in-progress` beats `done`), and what kind of knowledge it is. Results are taken in score order until the score falls below half of the best or the next one no longer fits the token budget, whichever comes first — at most two chunks per file.
 
 All of it lives in one SQLite file per project, `.capsa/index.db`. Every weight is a hypothesis; the `context_log` table is how they get corrected.