v0.9 release preparation

Retrieval infrastructure
that keeps up with AI.

Ketebe is an open-source retrieval and vector data platform for real-time ingestion, hybrid search, embedding lifecycle management, and governed AI-agent retrieval.

Start building → Read the docs →
Apache 2.0Rust-powered coreREST + gRPCFirst-party MCP
pythonbalanced retrieval
from ketebe import Client, QueryRequest

client = Client("http://localhost:7610")

result = client.query(
    "product-docs",
    QueryRequest(
        text="How does recovery work?",
        top_k=5,
        search_profile="balanced",
        explain=True,
    ),
)

# Dense + sparse + filters + reranking
✓ provenance included
✓ explainability enabled
collection: product-docstop_k 5
One retrieval planeDense · sparse · lexical · hybrid
Continuous dataDocuments · Kafka · async jobs
Model lifecycleEmbed · re-embed · rerank
Agent-nativeMCP · provenance · context budgets
Developer resources

Don’t just read the pitch.
Build with it.

Ketebe’s website now separates product positioning from the material you need to evaluate, integrate and operate the platform.

01 Docs Understand Ketebe

Concepts, retrieval, ingestion, embeddings, MCP, architecture, security and operations.

Explore → 02 Quickstart Run your first query

Create a collection, ingest a document and query it using the real Python SDK surface.

Explore → 03 Get started Choose your adoption path

Evaluate locally, pick an SDK, follow v0.9 packaging, or prepare for production.

Explore → 04 Roadmap See what ships next

v0.9 packaging milestones and the evidence-driven path toward a production-ready v1.0.

Explore → 05 Benchmarks Measure before claiming

The methodology Ketebe uses for reproducible latency, throughput, recovery and mixed-workload evidence.

Explore →
The platform

Less retrieval glue.
More product.

Modern AI systems often assemble a vector index, keyword engine, embedding workers, stream consumers, authorization proxies, and agent adapters as separate systems. Ketebe turns that fragmented path into one coherent retrieval platform.

Retrieve

Hybrid search without a sidecar maze

Dense, sparse, and lexical retrieval behind one query surface, with filters, fusion, reranking, and explainability.

↗
Ingest

Fresh data stays searchable

Ingest records and documents directly or continuously from Kafka. Ketebe owns the path into searchable state.

↗
Embed

Treat embeddings as lifecycle state

Server-side embedding and re-embedding workflows let models evolve without application-owned migration pipelines.

↗
Agent

A governed retrieval layer for agents

First-party MCP exposes discovery, retrieval, context assembly, provenance, and controlled ingestion.

↗
Operate

Durability is explicit

WAL-backed storage, recoverable segments, TLS/mTLS, backup, quotas, audit, and isolation are product contracts.

↗
Integrate

Use the interface your stack speaks

REST and gRPC plus first-party Rust, Python, TypeScript, Java, and Go SDKs.

↗
Architecture

Built around a recoverable source of truth.

Search performance matters. Recovery behavior matters more. Ketebe separates durable state from derived indexes so optimization never becomes the only copy of your data.

✓
Durable by design

The write path has an explicit source of truth. Derived indexes are rebuildable.

✓
Open and portable

Apache-2.0, self-hosted, API-first and designed without application lock-in.

✓
Built for changing models

Embedding and reranking can evolve without redefining your application data model.

✓
Governed retrieval

RBAC, quotas, audit and agent policy belong in the retrieval layer.

Explore architecture →
APPLICATIONS & AGENTS
PythonTypeScriptGoJavaRustMCP
KETEBE RETRIEVAL PLANE
DenseSparseFiltersFusionRerankExplain
IngestionDocuments · Kafka
EmbeddingLifecycle · Jobs
GovernanceRBAC · Quotas · Audit
Durable storageWAL · segments · recovery
source of truth
First-party MCP

Give agents retrieval.
Keep your boundaries.

Ketebe’s MCP adapter sits on the stable product API. Agents get discovery, multi-collection retrieval, reranking, provenance, context assembly and controlled ingestion—without a backdoor around authentication, authorization or storage correctness.

read-only by defaultproject scopedcontext budgeting
Read agent docs →
AI AGENTClaude · ChatGPT · IDE
tools
KETEBE MCPPolicy + Context
stable API
KETEBEGoverned Retrieval
Start small. Keep the architecture.

From zero to your first retrieval query.

The packaged v0.9 release is being finalized. You can already explore the real API and SDK model in the quickstart, then follow release readiness as the Docker Compose path lands.

01 Target v0.9 quickstart
$ docker compose up -d
ketebe-server  started
ketebe-data    ready

$ curl http://localhost:7610/health
{"status":"ok"}
Packaged v0.9 artifacts are being finalized.
01
Create a collection

Choose dimensions and similarity semantics.

02
Write records or documents

Bring vectors or let server-side embedding own the document path.

03
Query with one profile

Combine semantic and lexical signals with filters and reranking.

04
Connect agents and streams

Add MCP and Kafka without rebuilding the retrieval core.

Open quickstart →Choose another path →
Open source retrieval infrastructure

Build the knowledge layer
you can actually operate.

Read the docs, run the quickstart, inspect the roadmap and judge Ketebe by repeatable evidence.

Explore documentation →GitHub ↗