spectraltm-db

Pinecone-shape local vector database built on the Spectral encoding Rust engine.

spectraltm-db wraps the AVX2-accelerated spectraltm crate with a Pinecone-compatible Python API: same Index.create, Index.upsert, Index.query, metadata filters, namespaces, REST server — running entirely on disk, single-process, no cloud dependency.

Features

• Pinecone-shape surface: Index.create, open, upsert, query, fetch, delete, update, describe_index_stats, compact — wire-compatible with the pinecone Python SDK.
• Spectral encoding compression: 192 B/chunk at K=64, d=384 (vs ~1.5 KB for float32 embeddings); four named levels: spectral_k64 / k128 / k256 / float32.
• 11 metadata filter operators: $eq $ne $gt $gte $lt $lte $in $nin $exists $and $or — compiled to json_extract(...) SQL via a tiny in-process SQLite store.
• Multi-namespace by directory partitioning; per-namespace WAL, calibration buffer, raw.npy source-of-truth.
• Soft delete + real compact(): re-encodes from raw.npy minus tombstones, recovers on-disk space — not just a flag-flip.
• LangChain + LlamaIndex adapters (pip install spectraltm-db[langchain] / [llamaindex]).
• Pinecone-identical REST server (pip install spectraltm-db[server]): 11 endpoints, Api-Key auth, multi-index registry.
• CLI (python -m spectraltm_db {create,stats,upsert,query,delete,update,compact,serve,list}).
• AVX2 SIMD scoring via the Rust crate; falls back to scalar on other platforms.
• Optional SpectralIndex.score_one PyO3 binding for safe single-chunk scoring at n=1 (where select_nth_unstable_by would panic).

Installation

pip install spectraltm-db

Requires Python 3.10+. Pulls in spectraltm>=0.1.1 (Rust SSE engine) and a pure-Python stack: numpy, msgpack, stdlib sqlite3.

Optional extras:

pip install "spectraltm-db[encoder]"      # sentence-transformers auto-embedding
pip install "spectraltm-db[langchain]"    # langchain_core VectorStore adapter
pip install "spectraltm-db[llamaindex]"   # llama-index VectorStore adapter
pip install "spectraltm-db[server]"       # FastAPI REST server (11 endpoints)
pip install "spectraltm-db[all]"          # everything

Quick start

import spectraltm_db as stm

# Calibration sample: 1,000-10,000 random vectors of dim `dimension`.
# Calibrates the SSE encoder's magnitude/phase quantization grids.
calibration = [random.gauss(0, 1) for _ in range(2000 * 384)]

idx = stm.Index.create(
    name="conversations",
    path="~/.spectraltm/conversations",
    dimension=384,
    compression="spectral_k64",     # or spectral_k128 / spectral_k256 / float32
    calibration_sample=calibration,
)

# Upsert some vectors (Pinecone-style dicts)
idx.upsert(vectors=[
    {"id": "turn_001",
     "values": [...],                 # length-384 list of floats
     "metadata": {"session_id": "abc", "role": "user"}},
    {"id": "turn_002",
     "values": [...],
     "metadata": {"session_id": "abc", "role": "assistant"}},
])

# Pinecone-shape query
results = idx.query(
    vector=[...],                              # length-384 list
    top_k=10,
    filter={"session_id": {"$eq": "abc"}},     # metadata filter
)
for m in results["matches"]:
    print(m["id"], m["score"], m["metadata"])

# Reopen after process restart
idx.flush()
idx2 = stm.Index.open("~/.spectraltm/conversations")

See examples/rag_demo.py for a runnable end-to-end example.

Benchmarks — measured at v0.1.0.dev0 (median, top_k=10, dim=384)

Run via python -m spectraltm_db.examples.benchmark. Ingest uses the bulk add_batch Rust path; query latency is single-threaded.

Honest framing:

• Search latency scales linearly with N at AVX2 throughput (~12× slower for 10× more data). The n=1 corner case is handled by Rust score_one or NumPy fallback when the binding isn't present.
• Ingest is dominated by per-vector rFFT encoding. Each new vector pays an O(F log F) encoding pass; the bulk add_batch path amortizes Python→Rust overhead at 1.4 ms/vector but cannot skip the encoding. The fully-vectorized "load_codes" path (accepting pre-quantized codes) is on the v0.2 roadmap and is what unlocks billion-vector scale.
• spectraltm-db is a lossy compressed index, not a brute-force cosine replacement. See spectraltm's BEIR numbers (K=8 loses ~65% nDCG, K=64 ~12%) for the quality curve. Your choice of K is a storage-budget / quality-bar tradeoff.

When to use spectraltm-db

Use it when:

• You want a Pinecone-shaped API without a managed service: RAG, conversation memory, semantic search — anything you'd point at a Pinecone index and don't want cloud cost or egress.
• Your corpus fits on a single machine and on the order of N ≤ 1M vectors (above which encoding dominates ingest).
• You value single-tenant, on-disk persistence: no cluster, no replicas, just a folder on disk that you can cp, rsync, or back up however you like.

Don't use it when:

• You need real Pinecone parity in multi-region / serverless autoscaling / RBAC / backups — that's v0.2+.
• You need sparse-dense hybrid ranking with actual BM25/Reciprocal Rank Fusion — wire surface only in v0.1, scoring lands in v0.2.
• You need billion-vector scale — the AVX2 inner loop scales linearly with N; ingest cost dominates above ~1M without a pre-coded bulk-load path.
• You need multi-process writers — single-process at a time; multi-reader via SQLite WAL is fine; threaded HTTP works in-process.

Project layout

spectraltm_db/
├── spectraltm_db/
│   ├── __init__.py        # Index / IndexConfig / errors
│   ├── __main__.py        # CLI entry point
│   ├── config.py          # IndexConfig + CompressionSpec + MetricSpec
│   ├── errors.py          # exception hierarchy
│   ├── index.py           # Index — Pinecone-shape public API
│   ├── wal.py             # minimal JSONL write-ahead log
│   ├── sparse.py          # Sparse-dense wire surface (v0.2 stub)
│   ├── server.py          # FastAPI REST server (optional)
│   ├── .github/workflows/test.yml  # CI
│   ├── storage/
│   │   ├── vector_store.py     # wraps spectraltm.SpectralIndex
│   │   ├── id_map.py           # str-id ↔ chunk_id bijection
│   │   ├── metadata_store.py   # SQLite metadata + dynamic indexes
│   │   └── namespaces.py       # directory partitioning
│   ├── retrieval/
│   │   ├── filter_compiler.py  # dict → SQL
│   │   └── filter_executor.py  # SQL → chunk_id set
│   ├── integrations/
│   │   ├── langchain.py        # VectorStore adapter
│   │   └── llamaindex.py       # VectorStore adapter
│   └── examples/
│       ├── smoke_test.py
│       ├── rag_demo.py
│       └── benchmark.py
├── tests/                # 107 tests (3 conditional skips for optional Rust binding)
├── pyproject.toml        # pure-python; depends on spectraltm from PyPI
├── README.md
├── CHANGELOG.md
├── PUBLISH.md            # how to publish to PyPI
└── LICENSE               # Apache-2.0

How It Works

embeddings ─┐
            ├─► [encoder.calibrate]  ─► quantizer grids (mags, phases, norms)
            │
            ├─► [encoder.encode]      ─► SpectralCodes (idx, mag_q, phase_q, norm_q)
            │
            └─► [vector_store.add_batch] ─► Rust add_embeddings ─► dense (N,F) mags + phases

queries  ──┬─► [filter_compiler]   ─► SQL chunk_id whitelist (when filter present)
            │
            ├─► [encoder.encode]      ─► query SpectralCodes
            │
            └─► [SpectralIndex.search]  ─► AVX2 SIMD inner loop ─► top-K cosine scores
                                    │
                                    └─► [metadata_store.hydrate] ─► Pinecone-shaped response

The hot path is the Rust SpectralIndex.search (AVX2 SIMD on x86_64, scalar fallback elsewhere). Filter pre-narrowing happens in SQLite before we hand the candidate set to SSE; this matches the Pinecone data plane, where filter runs before scoring. n=1 uses an optional Rust score_one binding to bypass select_nth_unstable_by (which would panic for top_k >= n).

Development

# Editable install (Pure-Python) — pulls spectraltm from PyPI:
pip install -e ".[all]"

# Run the test suite:
pytest tests/ -v

# Reinstall the built wheel:
pip install --force-reinstall --no-deps dist/spectraltm_db-*-py3-none-any.whl

# Smoke CLI:
python -m spectraltm_db --help

For the optional score_one Rust binding path, you need a local build of spectraltm>=0.1.3 (workspace has the Rust patch; release to PyPI is on the v0.2 roadmap). Until then, VectorStore.search falls back to a NumPy dot product for the n=1 corner.

See PUBLISH.md for the PyPI release flow.

Citation

@misc{mckenzie2026spectraldb,
  title={spectraltm-db — Pinecone-shape local vector database on Spectral encoding},
  author={Mc Kenzie, Gerald Enrique Nelson},
  year={2026},
  month={6},
  day={29},
  howpublished={Companion to spectraltm (Sparse Spectral Encoding)},
  license={Apache-2.0},
  note={Uses spectraltm https://pypi.org/project/spectraltm/ as the underlying engine.}
}

Metadata

• Author: Gerald Enrique Nelson Mc Kenzie
• Date: 2026-06-29
• Repository: github.com/lordxmen2k/sparse-spectral-encoding
• Engine: spectraltm on PyPI
• License: Apache License 2.0
• Contact: lordxmen2k@gmail.com

License

Apache-2.0 — see LICENSE.

Built as a Pinecone-shape wrapper around the spectraltm crate. The two projects share zero source code per the explicit separation rule; the only cross-project coupling is the spectraltm PyPI dependency.