Python: Vector store filters raise NotImplementedError for negative number literals such as x.price > -5 #14571

What happens

A filter lambda that uses a negative number literal, such as lambda x: x.temperature > -5, raises NotImplementedError in nine of the eleven Python vector store connectors. Azure AI Search and Azure Cosmos DB NoSQL accept it.

Python parses -5 as UnaryOp(USub, Constant(5)), not as a constant. Each connector's _lambda_parser sends every UnaryOp other than not to a raise NotImplementedError(...), so the literal can never be translated. A negative bound is a normal thing to filter on (temperatures, scores, offsets, balances).

Where

python/semantic_kernel/connectors/ on current main (cc8a15f), the ast.UnaryOp branch of _lambda_parser:

• chroma.py line 409
• qdrant.py line 419
• mongodb.py line 494
• pinecone.py line 558
• weaviate.py line 456
• redis.py line 469
• postgres.py line 916
• sql_server.py line 644
• oracle.py line 1095

azure_ai_search.py and azure_cosmos_db.py handle USub by prefixing -, which is why x.price > -5 works there. The shared base class builds the AST in VectorSearch._build_filter (python/semantic_kernel/data/vector.py), so the literal can be normalized once for all connectors instead of in each parser.

Steps to reproduce

Real search against in-process Chroma:

import asyncio
from dataclasses import dataclass
from typing import Annotated

import chromadb

from semantic_kernel.connectors.chroma import ChromaCollection
from semantic_kernel.data.vector import VectorStoreField, vectorstoremodel


@vectorstoremodel
@dataclass
class Record:
    id: Annotated[str, VectorStoreField("key")]
    temp: Annotated[int, VectorStoreField("data", is_indexed=True)]
    vec: Annotated[list[float] | None, VectorStoreField("vector", dimensions=2)] = None


async def main():
    collection = ChromaCollection(record_type=Record, collection_name="temps", client=chromadb.EphemeralClient())
    await collection.ensure_collection_exists()
    await collection.upsert([Record("1", -10, [1.0, 0.0]), Record("2", -3, [0.0, 1.0]), Record("3", 4, [1.0, 1.0])])
    for f in ("lambda x: x.temp > -5", "lambda x: x.temp > 0"):
        try:
            results = await collection.search(vector=[1.0, 1.0], filter=f)
            print(f, sorted([r.record.id async for r in results.results]))
        except Exception as ex:
            print(f, type(ex).__name__, ex)


asyncio.run(main())

All connectors, calling the parser directly (collection._build_filter("lambda x: x.price > -5"), collection created with Cls.__new__(Cls) and only definition set):

Actual

chroma:    NotImplementedError: Unary +, -, ~ and ! are not supported in Chroma filters.
qdrant:    NotImplementedError: Unary +, -, ~ are not supported in Qdrant filters.
mongodb:   NotImplementedError: Unary +, -, ~ are not supported in MongoDB filters.
pinecone:  NotImplementedError: Unary +, -, ~ are not supported in Pinecone filters.
weaviate:  NotImplementedError: Unary +, -, ~, ! are not supported in Weaviate filters.
redis:     NotImplementedError: Unary +, -, ~ are not supported in RedisVL filters.
postgres:  NotImplementedError: Unary +, -, ~ are not supported in PostgreSQL filters.
sql_server:NotImplementedError: Unary +, -, ~ are not supported in SQL filters.
oracle:    NotImplementedError: Unsupported UnaryOp: <class 'ast.USub'>
azure_ai:  'price gt -5'
cosmos:    ('c.price > -5', [])

The Chroma script prints lambda x: x.temp > -5 VectorSearchExecutionException An error occurred during the search: Unary +, -, ~ and ! are not supported in Chroma filters. and then lambda x: x.temp > 0 ['3'].

Expected

x.temp > -5 is translated like any other number: {"temp": {"$gt": -5}} for Chroma, MongoDB and Pinecone, a Range(gt=-5) for Qdrant, "temp" > %s with [-5] for Postgres, and so on. The Chroma script returns ['2', '3'] for the first filter.

Suggested fix: in VectorSearch._build_filter, fold UnaryOp(USub, Constant(<int or float>)) into Constant(-value) before the parsers run. Other unary expressions (-x.field, -True, ~5, +5) are left to the parsers as they are today.

I checked open and closed issues and PRs for negative numbers, unary operators and filter parsing and found nothing covering this. I am opening a PR right after filing this.