Release notes

What changed in each version of dynavec, newest first.

dynavec follows semantic versioning. Upgrade with:

pip install --upgrade dynavec
uv pip install --upgrade dynavec

The full machine-readable history lives in CHANGELOG.md, and every version is a GitHub Release.

0.5.0 — 2026-09-16

More ingestion formats, more integrations, and observability.

Added

• Office document ingestion — DocxSource, PptxSource, and XlsxSource for Word, PowerPoint, and Excel files.
• Hugging Face Inference embedder — HFInferenceEmbedder backed by the HF Serverless Inference API.
• DSPy retrieval integration — DynavecRM(dspy.Retrieve) to back a DSPy pipeline with a dynavec client.
• Structured logging — opt-in JSON store-event logs with secret redaction (structured_logging=True).
• ProductQuantizer persistence — save() / load() via safe np.savez (no pickle).
• Dashboard dark mode with theme parity to the landing page.

Performance

• Vectorized MMR reranking — O(k·N) instead of O(k·N·k) on large candidate sets.

0.4.0 — 2026-09-12

Headlined by the in-memory hot tier for in-memory-engine latency without a paid cluster.

Added

• In-memory hot tier — hot_tier=True + db.warm(namespace) serve a namespace entirely from RAM (no S3 Vectors query, no DynamoDB hydration). Write-through keeps it current; filters/rescore/MMR run on the hot path; non-authoritative namespaces fall back to S3 safely. See In-memory hot tier.
• Retrieval quality runner — recall@k, MRR, and nDCG@k evaluation.
• Async LangChain retrieval — asimilarity_search and friends, so the retriever ainvoke() path runs on an owned async surface.
• Graph export to Mermaid and Graphviz DOT; an Ollama embedder; and URL + Markdown ingestion sources.

0.3.0 — 2026-09-10

A big feature release focused on more embedders, async, freshness, ingestion, and observability.

Added

• Async embeddings — aembed_documents / aembed_query on every backend, offloading sync clients to threads via asyncio.to_thread.
• New embedding backends — Mistral, Voyage AI, and Bedrock Titan multimodal image embeddings.
• SPFresh hot-tier — incremental hot-index rebalancing for freshly upserted vectors, so new data is searchable without a full rebuild.
• PDF ingestion source for the document pipeline.
• FastMCP server exposing semantic and graph search as MCP tools.
• Observability — a native telemetry recorder plus a stdlib dashboard on real query data, and a full Next.js + TypeScript + Tailwind + Recharts dashboard under dashboard/.
• max_pool_connections config, threaded into every boto3 client for high-concurrency workloads.
• list_vectors maintenance iterator, optional score normalization, cache hits/misses counters + stats(), an AWS doctor command, ingestion chunk deduplication, and a shipped py.typed marker.

Changed

• The semantic cache is now bounded by bytes rather than entry count.

Fixed

• Escape structured-storage key components to avoid namespace/id collisions.
• Drain QueryVectors pages fully and add a page_size control.
• Pin crewai away from the yanked 1.14.0 release.
• Resolve optional dependencies correctly on Python 3.9.

0.2.0 — 2026-08

Initial public release: the hybrid Amazon DynamoDB + Amazon S3 Vectors store, pluggable embedders, namespace RAG, product quantization, RRF fusion, MMR rerank, the GraphRAG layer, caching backends, framework adapters, and one-shot provisioning.