Run on Pixeltable 0.7.6, add Memory.search(), and add CI

[pierrebrunelle]

Why

Document, audio and video memories raised on construction against any Pixeltable >= 0.5.6. config.py carried the 0.4.2 iterator parameter names, and those names are not cosmetic — they go through dataclasses.asdict() straight into the iterator call, so each field name is an API contract:

pixelmemory passed	0.7.6 expects
chunk_duration_sec	duration
overlap_sec	overlap
min_chunk_duration_sec	min_segment_duration
html_skip_tags	skip_tags

The pixeltable.iterators shims forward **kwargs verbatim with no name mapping, which is why these were hard errors rather than deprecation warnings.

Separately, every retrieval example was dead code: they all opened with memory.chunk_views["..."], and chunk_views is a list of dataclasses on resources, not a mapping on Memory.

This went unnoticed because there is no CI on any branch and uv.lock pinned pixeltable 0.4.2. The most recent commit on main is 766c5ba "Removed test".

What is here

1. Green baseline + CI (d5091a3)

Iterator parameter renames and the deprecated shims swapped for document_splitter/audio_splitter/string_splitter/frame_iterator. Plus:

• the audio view output column is audio_segment, not audio_chunk
• transcriptions() kwargs were splatted top-level; the signature is (audio, model, model_kwargs), so WhisperParams(language='en') raised
• if_exists="replace_force" on five paths dropped and rebuilt every view and index on every construction, re-running Whisper and re-embedding everything
• non-nullable-by-default (0.7.3) made every field mandatory; Context gains required
• pxt.list_dirs() returns slash paths, so the dotted-namespace check never matched and nested namespaces failed on the second construction
• __getattr__ dereferenced self.resources from inside itself, so any __init__ failure became RecursionError
• vision.py off deprecated openai.vision; the anthropic path gains its required max_tokens and base64-encodes the image instead of passing a raw column reference as the data value

The suite is construction-only: no API key, no model weights, no spaCy, no media fixtures. conftest.py pops the provider keys from the environment so the tests prove it. It asserts on get_metadata()['iterator_call'], which pins the kwarg-rename class of bug precisely, and on view ids being stable across reconstruction, which pins replace_force. CI installs pixeltable unpinned with a weekly cron. Against the pre-fix source it reports 14 failures.

2. Memory.search() and Memory.views() (494ca94)

MemoryResources gains a search_targets list populated at each of the seven places an embedding index is created. Video registers its frames under "<col>_frames" so transcript and visuals can be searched separately.

memory.search("feline animal", limit=2)
memory.search(query, on="report")

search() always passes idx=, because chunked text carries two indexes on one column and a bare .similarity() there is ambiguous.

3. Repair the six broken examples (4ebcb9e)

Most collapse to one line. Two keep a raw handle deliberately and say why in a comment: the langchain example filters on metadata as well as ranking, and images.py returns the image alongside its caption. Two grep tests keep positional .similarity() and chunk_views[...] from coming back.

Public API

Memory(context=, namespace=, table_name=, if_exists=), memory.Entry/add(), the __getattr__ passthrough, and all Context field names are unchanged.

Deliberate breaks, both needing a 0.2.0 bump: the two splitter param dataclasses are renamed (unavoidable — those names are the iterator kwargs), and views/indexes now survive reconstruction instead of being rebuilt.

Still to do

Partial; landing it reviewable rather than sitting on it. Not yet done: uuid7() primary keys + Context(index=), as_tools()/lookup_tool(), Memory.router() replacing the 280-line FastAPI example, and examples/app.py.

A question worth your judgement

The audit raised whether the wrapper still earns its keep, and I think it partly does not. FastAPIRouter is verified working against runtime-built tables (add/search/lookup/delete all returned real rows via TestClient), which makes most of examples/fastapi/memory_service.py deletable. What survives is the one thing TableModel cannot express: a schema chosen at runtime from Memory(context=json.load(f)), plus the multimodal ingest pipelines. A dynamic TableModel is possible — I verified it — but it leans on sys._getframe(1) and a table-name-keyed registry, so I did not take it.

🤖 Generated with Claude Code