fix pgvector metadata filtering (parser + wire-up)

[Goodnight77]

description

Fix exact-match filter in pgvector parser and wire up filtered search

Previous version used json.dumps(value) for the filter value, which quotes strings with ", and Postgres reads double quotes as an identifier, not a string literal, so filtered queries still failed. Also, the parser was never called from search_one, so none of this ran

This PR:

• binds filter values as psycopg params (Jsonb(value)) instead of string-embedding them
• adds a payload column, stores record.metadata in it on upload
• calls the parser from search_one and applies the resulting WHERE clause

Tested against a real pgvector/pgvector:pg16 container exact-match and range filters return correct rows

[tellet-q]

Thanks for the PR!

Right now this code is a dead one, nothing executes it. So, the change is only required for possible future use. Apart from that it requires extension, alone it won't parse. json.dumps(value) renders a string as "Shoes", and double quotes in Postgres mean a quoted identifier, so you'd still get ERROR.

[Goodnight77]

@tellet-q
fixed both issues you raised:

• values now go through psycopg params instead of being string-embedded, no more double-quote/identifier bug
• wired it up for real: added a payload column, upload stores metadata in it, search now applies the filter

Tested against a real pgvector container, filters return correct results

[tellet-q]

This repo is mostly focused on benchmarking Qdrant, so pgvector engine is not the main line here. I still have a couple of notes for the change:

• Please move the tests from main to smth like tests/engine/clients/pgvector/test_pgvector_parser.py
• list-valued fields match nothing. payload -> 'labels' = '"lab_0"'::jsonb is false when labels is an array. It is an array in arxiv-titles-384-angular-filters. Qdrant (MatchValue) and Elasticsearch treat match-against-array as "contains", so pgvector would report precision 0 where every other engine scores normally.
• Unsupported filter types now abort the whole run. Before this PR search_one never called the parser, so nothing could raise, not applicable for most of the datasets, but still worth mentioning.
• Is GIN index really needed here?

[Goodnight77]

Thanks, @tellet-q all four addressed:

• tests moved to tests/engine/clients/pgvector/test_pgvector_parser.py
• exact match now uses @> instead of =: equality on scalars, "contains" on arrays, same as Qdrant/ES. Verified on PG 16, labels rows match
• geo datasets raise IncompatibilityError in configure, so they're skipped before upload
• GIN removed. It can't serve payload -> key expressions (checked with EXPLAIN), so it was dead weight