feat(logosdb): add LogosDB vector database integration

[jose-compu]

Summary

• Adds LogosDB as a supported vector database backend — a fast, embedded HNSW vector store written in C/C++ with Python bindings, backed by memory-mapped binary storage and hnswlib.
• Implements the full VectorDB interface: __init__, init context manager, insert_embeddings (via put_batch), search_embedding, and optimize.
• Registers DB.LogosDB in the enum and wires up init_cls, config_cls, and case_config_cls.
• Adds the logosdb CLI subcommand with a --uri flag (local directory path).
• Adds logosdb as an optional extra in pyproject.toml.

Design notes

• LogosDB is an embedded (single-process, file-based) database — no server required. The DB directory is passed via --uri.
• Distance metric is derived from the case MetricType at runtime (COSINE / L2 / IP). COSINE is the default and auto-normalizes vectors.
• Benchmark metadata IDs are stored as the text field (str(id)) and parsed back on search, since LogosDB's internal row IDs are independent of the benchmark ID space.
• HNSW index is built incrementally on insert; optimize() is a no-op with a log message.

Benchmark result

Tested on Performance1536D50K (OpenAI embeddings, 50K vectors, 1536 dim, COSINE) on Apple M-series:

Metric	Value
Load duration	340 s
Serial latency p99	4.6 ms
Serial latency p95	4.0 ms
Recall@100	0.9347
NDCG	0.9464

Test plan

• pip install logosdb (binary wheels for Linux x86_64/aarch64 and macOS x86_64/arm64, CPython 3.9-3.13)
• vectordbbench logosdb --uri /tmp/vdbbench_logosdb --case-type Performance1536D50K --skip-search-concurrent
• Verify recall, latency, and result JSON written to vectordb_bench/results/LogosDB/

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: jose-compu
To complete the pull request process, please assign xuanyang-cn after the PR has been reviewed.
You can assign the PR to them by writing /assign @xuanyang-cn in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[jose-compu]

can you please review @sre-ci-robot @jkatz @javiervegas @claude ?

[XuanYang-cn]

I found blockers in the current LogosDB integration. CI is red on the changed files, and the default command still enables a concurrent-search mode that conflicts with LogosDB's single-process database-path constraint.

[XuanYang-cn]

must-change: LogosDB inherits search_concurrent=True from CommonTypedDict, but LogosDB documents one DB directory as single-process while VDBBench concurrent search starts multiple ProcessPoolExecutor workers against the same --uri. The default command can fail or report invalid concurrent-search results after loading. Set parameters["search_concurrent"] = False or reject --search-concurrent for LogosDB until a supported single-process concurrent runner exists.

[jose-compu]

Thanks for catching this. Fixed in the latest commit by hard-setting parameters["search_concurrent"] = False in the CLI handler.

Quick note: I did test multi-process concurrent reads empirically (4 Pool workers opening the same DB path and running 50 searches each) and all succeeded without errors (LogosDB's memory-mapped storage appears safe for concurrent readers). That said, since the official docs declare it single-process, disabling concurrent search is the right conservative call for now. Can revisit if/when LogosDB formally documents multi-reader support.

Fixed here: b932872

[XuanYang-cn]

must-change: this os.path.exists() call fails the repo ruff gate with PTH110, so all PR test jobs are red before unit tests run. Replace it with Path(self.uri).exists() and update the imports.

[jose-compu]

ok thanks a lot, fixed here: 01e0fd0

[XuanYang-cn]

must-change: this import placement fails ruff I001, so CI stays red after adding the command. Run ruff check --fix vectordb_bench/cli/vectordbbench.py or move the import to the order ruff expects.

[jose-compu]

fixed here, possibly: 812eea6

[jose-compu]

@XuanYang-cn the benchmark is running now locally, can you try CI again please?

[jose-compu]

New results:

Processed from run 8bfdce6fe0a64bd8a43011bdbeb8c298 (2026-05-29). Serial search only — concurrent search was disabled for LogosDB.

Test configuration

Parameter	Value
Database	LogosDB
Case	Search Performance Test (50K Dataset, 1536 Dim)
Dataset	OpenAI-SMALL-50K
Vectors	50,000
Dimensions	1,536
Distance	Cosine
Top-K	100
Stages	drop_old → load → search_serial
DB path	/tmp/vectordbbench_logosdb
Run ID	8bfdce6fe0a64bd8a43011bdbeb8c298

Results

Metric	Value
Status	Success
Load duration	342.8 s (~5.7 min)
Insert duration	341.9 s
Optimize duration	0.9 s
Vectors loaded	50,000
Load throughput	~146 vectors/s
Recall@100	0.9347 (93.47%)
NDCG@100	0.9464 (94.64%)
Search queries	1,000
Search wall time	3.25 s
Serial QPS (derived)	~308
Mean latency	3.2 ms
P95 latency	4.0 ms
P99 latency	4.2 ms
Concurrent QPS	N/A (disabled)

Notes

• Summary qps=0.0 is expected: that field is for the concurrent-search stage, which LogosDB skips (search_concurrent=False).
• Serial QPS ≈ 1000 queries ÷ 3.25 s ≈ 308 QPS.
• Recall 93.5% and sub-5 ms P99 latency are solid for a 50K × 1536-dim cosine workload on embedded LogosDB.