Retune the llmwiki retrieval pipeline

CLI: llmwiki (vault /mnt/c/Users/emmaj/llmwiki, engines in .rag/). Retuning = llmwiki index (rebuild corpus parquet from vault markdown) then llmwiki tune (AutoRAG trial: BM25 + vector + hybrid-RRF race, scored on auto-generated QA, winner deployed).

When to call

• After /ingest lands new pages — until then they are search-visible but ask-invisible (ask retrieves from trial-time indexes).
• Staleness tell: llmwiki search X finds a page but llmwiki ask about it refuses — that's the signature; retune, don't debug retrieval.
• After changing .rag/config.yaml (nodes, prompt, models).
• NOT after mere page edits to already-indexed topics — BM25/embeddings tolerate content drift; batch retunes instead.

Procedure

1. llmwiki index — seconds; prints page count. Extra roots: llmwiki index wiki Clippings.
2. Corpus grew a lot or QA set should refresh? Delete .rag/data/qa.parquet first — tune reuses an existing QA set otherwise.
3. llmwiki tune — run backgrounded, 10–40 min on the 3090 (QA gen if needed + embed corpus + trial). Tell the user it's running.
4. On completion, sanity-check the leaderboard: read .rag/data/project/<latest>/*/*/summary.csv — expect hybrid ≥ BM25 ≥ vector on retrieval_recall, ROUGE ≈ 0.6. A collapsed metric (0.0) means a broken stage, not a bad corpus.
5. Verify end-to-end with one llmwiki ask about a recently added page.

Known failure modes (all hit in practice, all guarded — do not re-diagnose from scratch)

• QA gen crashes on batch 2 (Event ... bound to a different event loop): AutoRAG's legacy QA creator opens a new loop per cache batch while the Ollama client stays bound to the first. Guard: cache_batch ≥ content_size in .rag/engines/tune.py. If it recurs, that guard was lowered.
• doc_id ... not found in corpus_data from ask after tune: stale project-level corpus copy. tune.py deletes data/project/data/{corpus,qa}.parquet before each trial for exactly this; if seen, the deletion was removed.
• ollama is not a valid llm name: this AutoRAG build needs autorag.generator_models["ollama"] = Ollama registered before Evaluator/Runner — present in both tune.py and autorag_ask.py.
• node ... is not supported: this build uses lexical_retrieval/semantic_retrieval/hybrid_retrieval node types (docs still show the retired retrieval).
• Trials accumulate as data/project/0,1,2…; ask uses the highest. Old trials are safe to delete except the newest.