Introduction

Introduction

Veda is a next-generation multimodal vector database for the era of RAG. It stores, indexes, and searches embeddings from any modality — text, images, audio, video, and documents — and owns the entire path from raw files to fused, cross-modal search.

What makes Veda different

Most vector databases just store vectors: you bring your own embedding services, OCR pipeline, and chunking glue. Veda handles the whole pipeline natively, in a single pure-Rust engine.

  • Multimodal native — first-class support for text, image, audio, video, and PDF, with audio and video fully wired through the pipeline and automatic modality-based routing; point Veda at a unified omni embedding model and all modalities share one cross-modal space, so a text query retrieves audio, video, and images directly.
  • 11 vector index types — HNSW, DiskANN (Vamana graph), IVF-PQ, Flat, three RaBitQ variants, Sq4 + Sq8 scalar quantization, and SPANN (posting lists in object storage), so you can trade recall, latency and memory per collection. A suffix array for exact text patterns and scalar indexes for metadata filtering sit alongside them, plus a GPU CAGRA index on cuVS builds.
  • Multi-hop questions — the agent searches again when one lookup is not enough, joining facts across documents with no knowledge graph and no query planner.
  • Guardrails on retrieved documents — hostile instructions hiding inside your corpus are screened at ingest and dropped before an LLM sees them, at the tool-dispatch choke point that every agent surface shares.
  • Query Parquet in place — register a collection over existing Parquet files and search them with no ingestion and no second copy.
  • SIMD-optimized distances — SimSIMD kernels (AVX-512, ARM NEON, SVE) on by default, with an AVX2/scalar fallback, across 6 metrics (Cosine, Euclidean, DotProduct, Hamming, Manhattan, Jaccard).
  • Multi-vector fusion — RRF, WeightedSum, and Intersection strategies for cross-modal retrieval in a single API call.
  • Agents & MCP — a built-in agentic layer and an MCP server, so any MCP client (Claude Desktop, Cursor, VS Code) gets cross-modal retrieval over your data, secured by JWT auth and row-level security — with answer citations + faithfulness scoring because Veda owns the retriever.
  • Self-tuning RAG — Veda generates a labeled test set from your own corpus and auto-tunes the pipeline (chunker, index, fusion, rerank) for the best recall/faithfulness/latency, then applies the winner reversibly — optionally re-tuning every time the collection grows.
  • Built-in document intelligence — FlashParse OCR integration plus seven chunking strategies turn raw documents into searchable vectors with zero external tooling.
  • Self-hosted web RAG — crawl the open web into a collection and let agents search → fetch → ground over live pages, on open-source self-hosted services (crawl4ai + searxng) — no external API, no data leaving your trust boundary.
  • SQL-like filters — combine vector search with structured metadata predicates.
  • Blob storage — store the raw files alongside their embeddings via MinIO or any S3-compatible backend.
  • Lightweight by default — three mandatory services (Veda, MinIO, NATS); docker compose up and you have a vector database in under 30 seconds.

Beyond the basics

Veda also ships a distributed data plane (per-shard append-only logs with quorum replication, MVCC snapshot reads at ?at=<ts>, Percolator-style snapshot-isolation transactions, atomic reindex with alias cutover, cross-region WAN log shipping), JWT authentication with role-based access control and per-record row-level security, JSON-path filtering, record TTL with a cron scheduler, analytics aggregations, an OpenTelemetry-first observability stack with APM & performance observability (distributed tracing, profiling, GenAI/RAG traces with cost accounting, an index advisor), a vector-native AIOps layer (an automated Ops Advisor over 36 rules + node resource signals; semantic log categorization, spike attribution, incident correlation & novelty; change-point detection; capacity forecasting; and a cross-signal RCA bundle with proactive root-cause drafts), Log Insights, Service Reliability (SLOs, synthetic checks, AI triage), RAG & AIOps Intelligence (answer citations + faithfulness scoring, an eval harness, agent memory, streaming), and LLM Observability & Optimization (a semantic response cache, embedding-based guardrails, and a searchable trace store), the Niriksha visualization dashboard, and a Python SDK.

How these docs are organized

License

Veda is open source under the Apache License 2.0.