One database for enterprise AI

1. 01

   RAG is five services glued together

   Vector store, embedding service, chunking script, LLM gateway and a streaming layer: five processes, five deploys, five places for auth to leak.

   One binary does the pipeline. Ingest chunks and embeds server-side; POST /api/v1/ai/rag streams retrieved context and generated tokens over one SSE response, with citations.

   nebula-chunk · nebula-embed · nebula-llm · nebula-server
2. 02

   Vectors are an extension, not a native type

   Bolted-on vector types share a planner, buffer pool and vacuum designed for rows. Similarity search competes with the workload the engine was tuned for.

   HNSW is the core index. Cosine, L2 and dot metrics, int8 quantisation with dequantisation fused into the distance loop, and SQL that plans semantic_match() as a first-class retrieval step.

   nebula-vector · nebula-sql
3. 03

   Exact tokens get lost in embeddings

   Dense retrieval handles paraphrase but misses error codes, SKUs and function names. Teams run a second search engine to get them back.

   Hybrid search in the same index. Okapi BM25 and HNSW are fused with per-bucket weights, with optional query expansion and a pluggable reranker.

   nebula-bm25 · nebula-rerank · NEBULA_HYBRID_WEIGHTS
4. 04

   In-memory vector indexes vanish on restart

   A pod restart drops the index and someone re-ingests from source, paying for every embedding again.

   Durable by design. Every mutation is CRC32-framed and fsynced to the WAL before it is applied. Boot loads the newest snapshot and replays the tail; a torn frame is truncated, never trusted. Vectors are stored in the log, so recovery needs no embedder.

   nebula-wal · nebula-index/durability · nebula-backup
5. 05

   Analysts and agents cannot reach the vector store

   BI tools speak Postgres. Agents speak MCP. A bespoke REST dialect turns into adapters that nobody maintains.

   Every client speaks its own protocol. The same SQL engine serves REST, gRPC and the Postgres wire protocol, so psql works. A native MCP server gives agents 24 tools with auth passed through.

   nebula-pgwire · nebula-grpc · nebula-mcp · nebula-search
6. 06

   HA and multi-region are bolted on later

   Replicas fall behind silently, compaction deletes segments a follower still needs, and a write to the wrong node diverges forever.

   Replication is part of the design. Raft within a region, home-region XDCR across regions, follower ack cursors that clamp WAL compaction, a write guard that returns 409 on followers, and a region-aware client that retries on the right home.

   nebula-raft · nebula-cdc · nebula-client · operator/

• 01

  No garbage-collector pauses

  Memory is freed deterministically by ownership. Tail latency for search and RAG is not hostage to a collector, which matters when an agent waits on every call.

• 02

  Memory safety without a runtime

  Buffer overflows, use-after-free and data races are compile errors, not CVEs. The WAL and HNSW arena are written in safe Rust with checked UTF-8 chunk boundaries.

• 03

  Concurrency the compiler checks

  Tokio serves REST, gRPC, pgwire and SSE from one async runtime. The HNSW graph uses per-node, per-layer locks, so readers search while writers insert.

• 04

  Traits instead of glue

  Embedder, LlmClient, Chunker and Reranker are traits. Swap Ollama for OpenAI, Anthropic or Gemini by configuration; the deterministic mocks keep CI offline.

• SQL + JSON

  SQL over documents

  semantic_match() and vector_distance() beside filters, IN, array_contains, GROUP BY, inner joins and EXPLAIN. JSON metadata on every document.

  SQL and documents
• Vectors

  HNSW and hybrid search

  Native HNSW with three metrics and int8 quantisation, BM25 fusion, query expansion, reranking and per-bucket weights.

  Vector and hybrid search
• RAG

  Batteries-included RAG

  Smart chunking, embedding with a two-tier cache, grounded answers with [n] citations, history and caller-supplied context, streamed over SSE.

  RAG pipeline
• Agents

  MCP for AI agents

  24 tools across SQL, search, RAG, documents, per-user long-term memory, cluster health, replication, backups and audit. Every tool declares whether it writes.

  Agents and MCP
• HA

  Replicated, multi-region

  Raft within a region, home-region XDCR across regions, follower promotion, zero-downtime swap rebalance and a Kubernetes operator.

  High availability
• Operations

  Durable and observable

  WAL, snapshots, backup to S3 and restore, CDC, Prometheus metrics with Grafana dashboards, live log streaming, slow-query log, audit log and graceful degradation.

  Durability and operations

Read the architecture

git clone https://github.com/bwalia/nebuladb
cd nebuladb
./scripts/start.sh                 # tinyllama + nomic-embed-text
./scripts/start.sh --model llama3  # a larger chat model

docker pull bwalia/nebula-server:latest
docker pull bwalia/nebula-showcase:latest

helm install nebula deploy/helm/nebuladb

1. Ingest

   POST /api/v1/bucket/docs/document chunks and embeds server-side.

2. Search

   POST /api/v1/ai/search, vector or hybrid.

3. Answer

   POST /api/v1/ai/rag streams context then tokens.

4. Query

   psql -p 5433 with semantic_match().

• SQL subsetNo OR in WHERE, outer joins, subqueries, CTEs or HAVING. Unsupported syntax returns 400 sql_unsupported (SQLSTATE 42601 on pgwire).
• Simple query protocol onlypsql works; drivers that require the extended protocol (prepare / bind) do not yet.
• One writer per bucketMulti-region is active-active per cluster, with each bucket homed in one region. No multi-master writes to the same bucket.
• No cross-document transactionsEach upsert is its own durable commit.
• Coarse access controlAPI keys and JWT grant API access; there are no row- or field-level ACLs inside search results.
• Index in RAMThe working index lives in memory with snapshots on disk; there is no cold-storage tiering for rarely used vectors.