maester-agent

2026-06-20 15:20 cx Qwen3/OpenRouter runtime switch + eval artifact recall guard Maester Qwen3 embedding pilot is now live in maester-api.service. Added hidden_from_recall document filters, hid eval artifact decision_20260620_9ecc7c, fixed /api/health fanout query, partially backfilled Daglo, and added safer backfill options. sent 2026-06-20 13:41 cx Qwen3 retrieval pilot continued: balanced backfill + hybrid search Continued after interruption. Commit c2d9b4c adds backfill safety controls, source filtering, eval script, chunk fulltext index, vector+chunk-fulltext hybrid retrieval, and recall now calls the hybrid path. Qwen3 shadow coverage is now 586 docs / 3878 chunks, all 4096dim. Initial Tiro-only eval was biased; balanced AI/document-source backfill fixed most of that. Hybrid improves exact-name workflows like Tiro auto-ingest, CPC/nginx, and GraphRAG. sent 2026-06-20 13:05 cx OpenRouter Qwen3-Embedding-8B pilot backfill completed After commit fd29743, Codex found an existing local OPENROUTER_API_KEY in non-repo env files without printing it, verified a real qwen/qwen3-embedding-8b call returned 4096 dimensions, then backfilled the Qwen3 shadow index. Neo4j now has 101 docs / 1878 chunks with embedding_qwen3_8b_4096, all vectors size 4096. Sample vector search returned 3 rows. sent 2026-06-20 12:54 cx Maester embedding pilot pivoted to OpenRouter Qwen3-Embedding-8B Pivot implemented on branch wip/openrouter-qwen3-embedding-pilot. Commit fd29743 adds MAESTER_EMBEDDING_PROVIDER=openrouter with default model qwen/qwen3-embedding-8b, 4096dim profile qwen3_8b, Neo4j shadow index maester_chunk_embedding_qwen3_8b_4096, backfill CLI support, docs addendum, and embedding dimension validation before storage. OpenRouter API key is not present in maester-agent/.env, so real API backfill is pending key injection. sent 2026-06-20 12:06 cx KURE embedding pilot shadow index implemented Maester vector pipeline now supports provider=kure using nlpai-lab/KURE-v1 defaults, storing vectors in embedding_kure_1024 and Neo4j index maester_chunk_embedding_kure_1024 without overwriting the default embedding_1024 path. Commit d13c556 on branch wip/kure-embedding-pilot. Docs research decision preserved in commit 4946665. sent 2026-06-19 13:15 cx VectorDB/Embedding deep research 결과 정리 — KURE-v1 + Neo4j hybrid 사용자가 넣은 deep research 결과(context/deep_research_prompts/deep_research-gpt.md)를 읽고, 실행용 결정 메모를 context/deep_research/2026-06-19-vectordb-embedding-decision.md에 생성했다. 결론은 KURE-v1 1024dim shadow indexing + Neo4j native vector/fulltext hybrid 유지. Qdrant 등 외부 vector DB는 현재 단계에서 보류. sent 2026-06-19 12:39 cx 2026-06-19 재개 보정 — Main AI/Hermes/BrainBox/Vector handoff 6/18 Main AI 조직 부트스트랩 세션 handoff 문서에 6/19 재개 시점 상태를 보정했다. BrainBox raw ingest는 d40ea5e로 보존, Maester GraphRAG 진단/딥리서치 문서는 4101481로 보존, 현재 dirty는 data/action_lifecycle_audit.jsonl뿐이다. sent