feat(vector-store): add IBM Db2 AI Vector Search as a vector store provider

[DhruvChaturvediIBM]

Description

Adds IBM Db2 AI Vector Search as a fully supported vector store provider for Mem0 OSS. Db2 12.1.2+ ships a native VECTOR(n, FLOAT32) column type and VECTOR_DISTANCE function built into the engine — no extension, plugin, or sidecar process required. This means teams already running Db2 for their relational workloads can store and search memory embeddings in the same database, with no additional infrastructure.

Type of Change

• New feature (non-breaking change that adds functionality)

Breaking Changes

N/A

What's included

Core implementation

mem0/vector_stores/db2.py — Db2VectorStore implementing the full VectorStoreBase interface:

Method	What it does
insert	Bulk inserts via executemany, hashes IDs to CHAR(16), stores vectors with VECTOR(CAST(? AS CLOB), dim, FLOAT32) and metadata as SYSTOOLS.JSON2BSON
search	VECTOR_DISTANCE nearest-neighbour search with optional metadata filters and FETCH FIRST n ROWS ONLY
delete	Accepts both raw string IDs and pre-hashed CHAR(16) IDs
update	Can update embedding only, metadata only, or both in one call
get	Single-row lookup by ID, returns None for missing records
list	Full scan with optional metadata filters and top_k limit, returns the [[OutputData]] shape Mem0 expects
list_cols	Queries SYSCAT.TABLES for all user tables in the current schema
delete_col	DROP TABLE (no-ops safely when table is absent)
col_info	Returns {schema, table_name, row_count} from SYSCAT.TABLES
reset	Drop + recreate — table is fully usable immediately after

mem0/configs/vector_stores/db2.py — Pydantic Db2Config:

• Accepts either a pre-built ibm_db_dbi.Connection via client or raw connection_params (database, host, port, username, password, optional security/ssl_cert for TLS)
• Validates and normalises distance_strategy — supports EUCLIDEAN, COSINE, and DOT
• Configurable column names (text_field, id_field, metadata_field, embedding_field) so it can be pointed at an existing table schema
• Rejects unknown extra fields

Distance to similarity score conversion

Strategy	Formula	Range
EUCLIDEAN	1 / (1 + distance)	(0, 1]
COSINE	max(0, 1 - distance)	[0, 1]
DOT	returned as-is	unbounded

Metadata filtering

Filters run against the SYSTOOLS.BSON2JSON metadata column via JSON_VALUE. Supported filter shapes:

• Scalar equality — {"user_id": "alice"}
• List / IN — {"category": ["work", "personal"]}
• Wildcard existence — {"run_id": "*"} (skipped in WHERE, matches any row)
• Multi-field AND — top-level keys are combined with AND

Registration

• mem0/vector_stores/configs.py — "db2" added to the provider registry so VectorStoreConfig(provider="db2", config={...}) and Memory.from_config({"vector_store": {"provider": "db2", ...}}) both work
• pyproject.toml — ibm_db>=3.2.0 added to the vector-stores optional dependency group

Tests

tests/vector_stores/test_db2.py — 109 tests total:

• 93 unit tests — fully mocked (ibm_db_dbi is stubbed via sys.modules when not installed), cover config validation, ID hashing, distance-to-score conversion, table creation, every CRUD method, NULL handling, filter SQL generation, and VectorStoreConfig provider registration
• 15 live integration tests — run against a real Db2 instance, guarded by a @requires_db2_credentials skip marker, cover the full CRUD lifecycle plus test_live_documentation (end-to-end via Memory.from_config)
• Credentials loaded automatically from examples/misc/.env via python-dotenv — no manual export needed

Documentation

• docs/components/vectordbs/dbs/db2.mdx — new provider page with requirements, Memory.from_config usage example, full config table, score conversion explanation, and metadata filter examples
• docs/docs.json — db2 added to the Supported Vector Databases navigation group
• docs/llms.txt — Db2 entry added so the docs-llms-txt-check CI check stays green

Example

examples/misc/db2_vector_store_example.py — a self-contained runnable demo (no LLM needed) that walks through all 10 operations against a live Db2 instance: connect, insert, similarity search, filtered search, list with filter, get by ID, update, delete, collection info, and reset.

Test Coverage

• I added/updated unit tests
• I added/updated integration tests
• I tested manually — all 108 tests pass against a real Db2 12.1.2 instance; the example script runs end-to-end successfully

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my code
• I have added tests that prove my fix/feature works
• New and existing tests pass locally
• I have updated documentation if needed