[TrinaxCode]

Description

Experiment, draft: do not merge until the measurement below is filled in.

Every chunk is embedded together with nine lines of identical metadata (ids, absolute root, collection, pipeline version). On short chunks that boilerplate can blur the vector and make retrieval pick the wrong chunk. This change embeds only rel_path and file_name; the rest of the metadata stays on the node for filtering and citations. extractor_version goes to 3 so existing indexes rebuild.

Measurement (code Q&A, 4 TrinaxAI source files + 1 unanswerable question, real qwen3-embedding:0.6b and qwen3.5:4b)

	recall@5	MRR	nDCG	citation precision	answer terms	no-answer correct
baseline (main)	1.0	0.7	0.926	0.6	1.0	1.0
this change	pending	pending	pending	pending	pending	pending

I will fill this in, and close the PR if it does not improve on the baseline.

Type of change

• Bug fix
• Other: retrieval-quality experiment

Checklist

• Index tests pass locally; new test covers the embedded text.
• Measured improvement (pending).

---

Generated by Claude Code

[TrinaxCode]

Measured and rejected. Same corpus (4 TrinaxAI source files, 37 chunks, 5 questions), real qwen3-embedding:0.6b and qwen3.5:4b, same questions:

	recall@5	MRR	nDCG	citation precision	answer terms	no-answer correct
baseline (main)	1.0	0.7	0.926	0.6	1.0	1.0
embed only path + file name	1.0	0.7	0.926	0.6	0.8	1.0

Retrieval is identical, so the repeated metadata is not what limits ranking on this corpus; the one-case drop in answer terms is within the noise of a single generation by a 4B model. The 0.6 citation precision comes from how many files get cited, not from the vectors. Closing without merging; no change to main. (Limitation: 37 chunks is a small corpus, so this does not rule out an effect on a large one.)

---

Generated by Claude Code