fix(searcher): use filter-fallback in CLI search path for HNSW drift (#1665)

[rodboev]

Summary

search() (CLI path) and search_memories() (MCP/API path) diverged in their handling of filtered HNSW failures. search_memories already routes through _query_drawers_with_filter_fallback (added for #1245/#1035), which retries unfiltered and post-filters in Python when ChromaDB raises Error finding id on a where=-scoped query. search() did not — its except Exception block at searcher.py:381 caught the error and re-raised as SearchError, so wing-filtered CLI searches failed with a cryptic message while the equivalent MCP call recovered cleanly.

The fix routes search() through the same helper, making CLI and MCP behavior consistent. No new infrastructure: the helper already handles the unfiltered-retry, Python post-filter, and warning log.

Fix

Replace the inline col.query(**kwargs) call in search() (searcher.py:379) with _query_drawers_with_filter_fallback(col, kwargs, query, n_results, wing, room). The helper signature accepts the same kwargs dict and the existing wing/room locals, so no refactoring is needed around the call site.

Add a regression in tests/test_searcher.py: mock a collection whose query raises only when where= is present, call search() with a wing filter, assert no SearchError is raised and results are returned.

Related work

Open PRs #951, #1005, #1463, and #1598 address overlapping search-fallback surfaces. This PR is the narrowest fix: it routes the CLI search() path through the existing _query_drawers_with_filter_fallback helper that search_memories() already uses. #1005 and #1598 both subsume this change as part of broader scope; if either lands first, this PR is redundant.

Test plan

• python -m pytest tests/test_searcher.py -v — 31 passed
• ruff check . — clean
• ruff format --check . — clean

[gemini-code-assist]

Code Review

This pull request updates the search mechanism in mempalace/searcher.py to use a fallback query function (_query_drawers_with_filter_fallback) when a filtered query fails, allowing it to fall back to an unfiltered query and post-filter the results in Python. A corresponding unit test has been added in tests/test_searcher.py to verify this behavior. I have no feedback to provide as there are no review comments.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[mvalentsev]

The fix works, but check the #1665 thread before review: four open PRs already cover this same failure (#1005 delegates CLI search to search_memories and removes the CLI/API split entirely, #1598 hardens CLI and API together, #951 carries a get()-based variant, #1463 a BM25 fallback). This makes a seventh remedy on the cluster, and the description mentions none of them. Choosing between copying the fallback into the second path and collapsing the two paths is a maintainer call.

[rodboev]

Fair point. I should have acknowledged those PRs in the description. Updated it to reference #951, #1005, #1463, and #1598 and note the overlap. This one is the narrowest surgical option for the CLI-only gap; if any of the broader PRs land first it's redundant.

[mvalentsev]

@rodboev Just one request, please verify problem/fix by actually running MemPalace, rather than reviewing the code theoretically.

[rodboev]

Ran it locally against a small test palace. CLI search works fine with and without wing filters on a healthy index. The tricky part is reproducing the actual drift scenario from #1665, since it requires orphaned HNSW IDs left behind after a bulk delete. The fix itself is straightforward routing: search() now goes through the same _query_drawers_with_filter_fallback helper that search_memories() already uses, so the fallback logic is shared instead of duplicated. I can set up a targeted repro if you'd like, but it would need a palace large enough for the bulk-delete path to leave stale IDs in the index.

[rodboev]

Withdrawing this PR. The batch had serious quality problems, as you rightly pointed out. I genuinely did not see the issues and apologize for that. Closing everything and going back to the drawing board. Sorry for the noise.