Pinecone diagnosis: recall worse than expected, isolate model vs index

Diagnosis

Symptom

Queries return plausible but incomplete results; known-good neighbors are missing from matches.

Likely cause

Either the embedding model does not capture the similarity you want, or the index's approximate search is dropping candidates. These need different fixes, so do not guess.

Confirm it

Follow the docs' recall-debug procedure:

1. Build an eval set: at least 10 queries over a 100k-vector source set, with labeled right answers.
2. Embed queries and corpus with your model; run brute-force search (FAISS or numpy) and score against the labels. If brute force scores badly, it is a model problem.
3. Upsert the same vectors to Pinecone, run the same queries, and compare Pinecone's returned ids to the brute-force ids. Pinecone recall near 0.99 on s1/p1 means the index is fine and the model is the issue.

Fix

1. Model problem: change the embedding model (and rebuild the index at the new dimension).
2. Index problem: check top_k (too small clips recall), filters (too strict), and pod size; escalate to support with the reproducible dataset if recall stays low.

Verify the fix

Re-run the eval set after the change and compare the recall number, not vibes. Keep the eval set; re-run it on every model or index change.

Pinecone's debug-model-vs-pinecone-recall troubleshooting page prescribes this exact six-step procedure: eval framework, brute-force baseline, Pinecone comparison, ~0.99 expected recall on s1/p1, escalate with reproducible data. Verified against live docs 2026-09-26.