Vector search at storage prices

S3 is the source of truth. Nodes are stateless. Vector, full-text, and hybrid search.
Fully open source.

Install the Zeppelin skill for Claude Code.
Use our hosted instance to create namespaces, store and query data!
Just ask Claude to tell you more about this skill and try everything out. 

curl -sL https://zepdb.github.io/install.sh | sh
Storage

S3-Native Storage

Object storage is the source of truth. Scales to petabytes.

Search

Vector Search

IVF-Flat with SQ8, Product Quantization, and hierarchical ANN.

Full-Text

BM25 Full-Text

Inverted indexes with tokenization, stemming, and hybrid ranking.

Filters

Bitmap Pre-Filters

RoaringBitmap indexes for sub-ms attribute filtering.

Consistency

Tunable Consistency

Strong or eventual reads. Choose per query.

Verification

Formally Verified

26 TLA+ specs prove WAL, compaction, and CAS correctness.

451 tests
26 TLA+ specs
1.2M+ states explored
Apache-2.0

Performance at scale

1M vectors, 768d, cosine

zepdb/zeppelin-benchmark
Vector Search
Warm p50
Zeppelin
~45ms
Turbopuffer
~8ms
Warm p99
Zeppelin
~120ms
Turbopuffer
~35ms
Cold
Zeppelin
~800ms
Turbopuffer
~343ms
Full-Text Search
Warm p50
Zeppelin
~20ms
Turbopuffer
~11ms
Cold
Zeppelin
~120ms
Turbopuffer
~221ms

Pricing calculator

Zeppelin = S3 storage + compute. Configure your setup and compare.

10M
Monthly cost
Storage (S3) $0
Compute $0
Total $0/mo

SQ8 quantization, 20% index overhead, 730 hrs/mo compute

Cost comparison

Up and running in 60 seconds

1

Start

docker run -p 3000:3000 \
  -e AWS_ACCESS_KEY_ID=your-key \
  -e AWS_SECRET_ACCESS_KEY=your-secret \
  -e ZEPPELIN_BUCKET=your-bucket \
  -e ZEPPELIN_REGION=us-east-1 \
  ghcr.io/zepdb/zeppelin:latest
2

Create a namespace

curl -X POST http://localhost:3000/v1/namespaces \
  -H "Content-Type: application/json" \
  -d '{"dimensions": 768}'

# Response includes "name": "<uuid>" — save it!
3

Upsert

curl -X POST http://localhost:3000/v1/namespaces/<namespace-uuid>/vectors \
  -H "Content-Type: application/json" \
  -d '{
    "vectors": [
      {
        "id": "vec-1",
        "values": [0.1, 0.2, 0.3, "..."],
        "attributes": {
          "category": "science",
          "year": 2024
        }
      }
    ]
  }'
4

Query

curl -X POST http://localhost:3000/v1/namespaces/<namespace-uuid>/query \
  -H "Content-Type: application/json" \
  -d '{
    "vector": [0.1, 0.2, 0.3, "..."],
    "top_k": 10,
    "filters": {
      "op": "And",
      "conditions": [
        {"op": "Eq", "field": "category", "value": "science"},
        {"op": "Gte", "field": "year", "value": 2020}
      ]
    },
    "include_attributes": true
  }'

Architecture

Stateless nodes. S3 is the single source of truth.

Zeppelin architecture diagram
Write Path
  • Append-only WAL fragments to S3
  • OCC via S3 ETags on central Manifest
  • Monotonic fencing tokens prevent zombie writers
  • Deterministic sequence numbers (no clock skew)
Read Path
  • Merges fresh WAL data + optimized segments
  • Strong: parallel scan of WAL + indexed segments
  • Eventual: fast-path over segments only
  • Singleflight coalesces concurrent manifest fetches
Compaction
  • IVF-Flat / Hierarchical ANN with k-means + quantization
  • BM25 inverted indexes for full-text search
  • RoaringBitmap indexes for sub-ms attribute filtering
  • Runs in background, never blocks writes
Benefits
  • Instant elasticity: stateless nodes, no resharding, 1→100 in seconds
  • Zero-state recovery: no replication lag, Manifest is source of truth
  • Storage economics: S3 rates vs. NVMe; scale compute independently
  • Decoupled read/write: WAL available immediately; indexing is background
Tradeoffs
  • S3 latency floor: baseline 20-50ms; caching helps but not sub-5ms
  • Consistency lag: ManifestCache TTL ~500ms across nodes
  • Compaction health: latency degrades if WAL fragments grow unchecked
  • Cold-start: first queries populate LRU cache from S3

About

Anup Ghatage

Anup Ghatage

Founder, Zeppelin

  • 14 years in databases and distributed storage
  • Principal SWE, Salesforce
  • Apache BookKeeper committer
  • SAP HANA · Cisco WebEx · Sybase ASE
  • 7 patents · 10+ Hackathon wins
  • Carnegie Mellon · University of Pune (CS)

Your vectors belong in your object storage, not someone else's proprietary system. Search should cost what storage costs.

I've spent 14 years inside distributed storage systems: Sybase ASE internals in C/C++, Cisco WebEx, SAP HANA, then building Hyperforce at Salesforce. One lesson kept coming back.

In an agent-first future, agents won't care about latency the way web apps do. They'll care about cost and simplicity. If your vectors live in S3, you pay S3 prices. Nodes become stateless. You scale by adding buckets, not by babysitting clusters.

On February 10th, 2026, I started building Zeppelin at the Claude Code Hackathon. The seed fund was $500 in Claude credits. I directed the architecture. Every line of Rust was written by Claude Opus 4.6. One human, one AI, no team, no investors.

The result: 451 tests, 26 TLA+ specs, 1.2M+ states explored, and an engine released under Apache-2.0.

Zeppelin is open source because infrastructure should be. No vendor lock-in, no opaque pricing, no proprietary formats. Your data stays in your buckets.

Smarter architecture beats faster clusters.