---
title: "Qwen3 retrieval pilot continued: balanced backfill + hybrid search"
from: codex
to: ["claude-code", "jadong"]
project: "maester-agent"
date: 2026-06-20T13:41:31.381440+09:00
status: sent
type: result
---

## Result

- Branch: `wip/openrouter-qwen3-embedding-pilot`
- New commit: `c2d9b4c feat(retrieval): add balanced Qwen backfill and hybrid eval`
- Qwen3 coverage after balanced backfill: `586` docs / `3878` chunks, min/max dim `4096`.
- Added backfill controls: `--skip-existing`, `--source`, `--exclude-source`, `--max-chunks-per-doc`, `--max-total-chunks`.
- Added chunk fulltext index: `maester_chunk_fulltext`.
- Added vector+chunk-fulltext hybrid search and wired Maester recall to use it.
- Added eval script: `scripts/eval_vector_retrieval.py`.

## Eval Files

- `context/evaluations/2026-06-20-qwen3-vector-pilot.json`
- `context/evaluations/2026-06-20-qwen3-vector-pilot-balanced.json`
- `context/evaluations/2026-06-20-qwen3-hybrid-pilot-balanced.json`

## Observations

- Tiro-only pilot was misleading: generic/system queries all returned recent Tiro transcripts.
- Balanced AI/document-source backfill shifted top docs to Claude/Gemini/Codex/design docs.
- Hybrid retrieval improved exact-name workflows for `Tiro`, `CPC/nginx`, and `GraphRAG`.
- Remaining gap: current repo-local Qwen/OpenRouter pivot docs must be ingested into Maester too; otherwise recall cannot retrieve them as Maester documents.

## Dirty State

- `data/action_lifecycle_audit.jsonl` remains dirty as unrelated runtime audit log and was not committed.