The agent stack in your AWS account
Open-source, serverless memory, orchestration, observability and evaluation — one Python library.
dynavec starts as a hybrid vector database that fuses Amazon DynamoDB (single-digit-millisecond document store) with Amazon S3 Vectors (billion-scale serverless ANN), and grows into agents, traces and evals. Everything runs in your AWS account and bills only when you use it — no external control plane.
Agent lifecycle
Remember · dynavec
Store and search memory: hybrid BM25 + ANN search, GraphRAG, hot tier and caching.
Get started › In progressBuild · dynaflow
Compose agents: chat models, ReAct and planner, tool registry with MCP, supervisor routing.
Get started › EarlyObserve · dynalogs
Structured per-run and per-node logs with secret redaction, plus a telemetry dashboard.
Get started › PartialMeasure · dynaevals
Score retrieval and answers with recall@k, MRR, nDCG and LLM-judged quality.
Get started ›Start here
Installation
pip or uv, plus optional embedder and framework extras.
Install ›Quickstart
Your first upsert and semantic search in a few lines.
Start ›Configuration
Every DynavecConfig option, with sensible defaults.
Configure ›Guides
Embeddings
Pluggable, bring-your-own-key: OpenAI, Gemini, Cohere, Bedrock, local.
Search
Semantic search with metadata pre-filtering and document hydration.
Metrics & rerank
Cosine, dot, euclidean, manhattan, weighted combos, and MMR.
Namespaces
Per-tenant / per-collection isolation on one index.
Knowledge graph
Entity-relationship traversal that guides the vector search.
Caching
Semantic, DynamoDB-TTL, or Redis / ElastiCache query caches.
Product quantization
Compress cached vectors up to 32× with ADC distance.
Ingestion & MCP
Pull, chunk, embed from any source — including any MCP server.
Credentials & IAM
Access keys, profiles, cross-account assume-role, least-privilege policy.
Integrations
LangChain, LlamaIndex, and a tool for LangGraph / CrewAI / Strands.
FAQ
Common questions about regions, limits, consistency, and costs.
Benchmarking
Recall, latency, and cost modeled across dimensions and scale.
Ecosystem
The four pillars — dynavec, dynaflow, dynalogs, dynaevals — and the roadmap.