[integrations] Fix enhanced-mcp search returning zero rows and captures losing embeddings

[TomKeyser]

Contribution Type

• Integration (/integrations)

What does this do?

Fixes three defects that make integrations/enhanced-mcp non-functional against the schemas it ships against. All search returned zero rows, captures were stored with a NULL embedding (and reported an error while doing so), and get_thought could not be called at all. All three were found by testing a fresh install built from docs/01-getting-started.md plus schemas/enhanced-thoughts.

1. All search returned zero rows

brain_search_thoughts (both semantic and text modes) and search_thoughts_text passed {exclude_restricted: true} — plus start_date/end_date — inside the RPC filter payload. But both RPCs treat that payload as a metadata containment test:

AND t.metadata @> coalesce(p_filter, '{}'::jsonb)

So the query asked Postgres for thoughts whose metadata literally contains {"exclude_restricted": true}. No thought carries that key, so every semantic and full-text query matched nothing. A comment in the code assumed older RPC versions would ignore unknown filter keys — they don't; a non-matching containment test returns zero rows.

Fix: pass only genuine metadata filters, and apply sensitivity and date filtering client-side. The semantic branch already post-filtered sensitivity_tier !== "restricted", which made the payload flag both redundant and fatal; the text paths now do the same.

2. Captures were stored with a NULL embedding

brain_capture_thought sends the embedding in p_payload, but the upsert_thought in schemas/enhanced-thoughts never reads it — embedding is not in that function's INSERT list. The handler then checked result.thought_id, while that RPC returns {id, fingerprint}, so it threw "upsert_thought returned no result" after the row had already been inserted.

The result was an error message on a write that actually succeeded, leaving a thought with a NULL embedding that semantic search can never return. Re-capturing does not heal it, because the fingerprint dedup path short-circuits before computing an embedding.

Fix: accept either key, then persist the embedding in a follow-up update — the same two-step the stock core server in server/index.ts already performs — and report explicitly if that update fails rather than claiming success.

3. get_thought could not be called

It declared id as z.number().int().min(1), but the core schema defines thoughts.id as uuid default gen_random_uuid(). No valid call was possible on a standard install. The lookup is a plain .eq(), so accepting a uuid string is sufficient.

Deliberately not fixed here

update_thought and related_thoughts carry the same integer-id declaration. related_thoughts additionally requires the knowledge-graph schema, which I don't have installed, so I couldn't verify a change to either and left them alone rather than ship untested edits. Happy to follow up if a maintainer with that schema can confirm the expected id type.

Separately, graph_search throws a raw Could not find the table 'public.entities' rather than the graceful "install required schema" message the README describes. Left as-is to keep this PR scoped.

Requirements

No new requirements. Same services as the existing integration: Supabase (with pgvector) and OpenRouter. No dependency, schema, or API changes — this is a behavior fix confined to integrations/enhanced-mcp/index.ts.

Testing

Tested on my own Open Brain instance (Supabase + pgvector, us-west-2, Postgres 17), built from docs/01-getting-started.md with schemas/enhanced-thoughts applied.

Before:

• search_thoughts_text → {"results":[]} for every query, including exact phrases present in stored content
• brain_search_thoughts → [] in both text and semantic mode
• brain_capture_thought → Error: upsert_thought returned no result (row inserted anyway, embedding NULL)
• get_thought → uncallable; schema rejects a uuid

After:

• search_thoughts_text returns a phrase that semantic search scored at 0.348 and silently missed (rank 0.6)
• brain_search_thoughts text mode returns both stored thoughts, ranked 0.665 / 0.6
• brain_search_thoughts semantic mode returns correctly ordered results at 0.507 / 0.481
• brain_capture_thought returns {thought_id, action: "inserted"} and the thought is retrievable semantically at 0.590
• get_thought returns full metadata and provenance for a uuid
• count_thoughts and brain_thought_stats unaffected, still correct

Checklist

• I've read CONTRIBUTING.md
• My contribution has a README.md with prerequisites, step-by-step instructions, and expected outcome (existing integration README unchanged — this is a bug fix, no doc changes needed)
• My metadata.json has all required fields (unchanged)
• If my contribution depends on a skill or primitive, I declared it in metadata.json and linked it in the README (no new dependencies)
• I tested this on my own Open Brain instance
• No credentials, API keys, or secrets are included

[github-actions]

Hey @TomKeyser — welcome to Open Brain Source! 👋

Thanks for submitting your first PR. The automated review will run shortly and check things like metadata, folder structure, and README completeness. If anything needs fixing, the review comment will tell you exactly what.

Once the automated checks pass, a human admin will review for quality and clarity. Expect a response within a few days.

If you have questions, check out CONTRIBUTING.md or open an issue.