Files
skills/public/hanzo-vector.md

19 KiB

Hanzo Vector - High-Performance Vector Search Engine

Category: Hanzo Ecosystem Related Skills: hanzo/hanzo-database.md, hanzo/hanzo-storage.md, hanzo/hanzo-llm-gateway.md

Overview

Hanzo Vector is a high-performance vector search engine for the Hanzo AI platform. It is a full fork of Qdrant v1.17.0, written in Rust, providing dense and sparse vector similarity search with filtering, payloads, quantization, distributed clustering, and GPU acceleration.

The crate is named hanzo-vector (package name in Cargo.toml), but the binary is still qdrant for wire-protocol compatibility with Qdrant clients.

Why Hanzo Vector?

  • Drop-in Qdrant replacement: Compatible with all Qdrant client SDKs (Python, JS, Rust, Go)
  • Dense + sparse vectors: Hybrid search combining semantic embeddings and keyword matching
  • HNSW indexing: Approximate nearest neighbor with SIMD acceleration
  • Vector quantization: Scalar, product, and binary quantization for memory efficiency
  • GPU acceleration: NVIDIA (Vulkan) and AMD (ROCm) support via --features gpu
  • Distributed mode: Raft consensus, sharding, and replication
  • REST + gRPC APIs: OpenAPI 3.0 schema and Protocol Buffer definitions
  • Write-ahead logging: Crash-safe persistence with snapshot-based backup
  • Edge runtime: Embedded mode via lib/edge with Python bindings

Tech Stack

  • Language: Rust (edition 2024, rust-version 1.92)
  • HTTP server: actix-web with rustls TLS
  • gRPC server: tonic (custom fork for qdrant compatibility)
  • Consensus: Raft (via raft crate, prost codec)
  • Allocator: jemalloc (tikv-jemallocator) on x86_64/aarch64
  • Profiling: Pyroscope, tracing-tracy, console-subscriber
  • Build: cargo-chef for Docker layer caching, mold linker

OSS Base

Fork of qdrant/qdrant. Repo: hanzoai/vector, branch master (not main).

When to use

  • Storing and searching AI embeddings (LLM, vision, audio)
  • Semantic similarity search with metadata filtering
  • Recommendation engines using vector similarity
  • RAG (Retrieval-Augmented Generation) pipelines
  • Hybrid search combining dense vectors with sparse (BM25-style) vectors
  • High-throughput vector operations requiring GPU acceleration

Hard requirements

  1. Rust 1.92+ for building from source
  2. Protobuf compiler (protoc) for gRPC code generation
  3. clang + lld for linking (or mold)
  4. Docker for containerized deployment
  5. Vulkan drivers if using GPU features (NVIDIA or AMD)

Quick reference

Item Value
Repo github.com/hanzoai/vector
Branch master (code), main (stub placeholder)
Version 1.17.0
Language Rust (edition 2024)
Binary qdrant
REST port 6333
gRPC port 6334
P2P port 6335 (cluster mode)
Docker image ghcr.io/hanzoai/vector:latest
License Apache-2.0
Config config/config.yaml
OpenAPI openapi/
gRPC protos lib/api/src/grpc/proto/

One-file quickstart

Docker (single node)

docker run -p 6333:6333 -p 6334:6334 ghcr.io/hanzoai/vector:latest

Docker with persistent storage

docker run -p 6333:6333 -p 6334:6334 \
 -v $(pwd)/data:/qdrant/storage \
 -v $(pwd)/snapshots:/qdrant/snapshots \
 ghcr.io/hanzoai/vector:latest

Docker with custom config

docker run -p 6333:6333 -p 6334:6334 \
 -v $(pwd)/config.yaml:/qdrant/config/production.yaml \
 -v $(pwd)/data:/qdrant/storage \
 ghcr.io/hanzoai/vector:latest

Build from source

cargo build --release --bin qdrant
./target/release/qdrant --config-path config/config.yaml

Build with GPU support

cargo build --release --bin qdrant --features gpu

Core Concepts

Architecture

┌──────────────────────────────────────────────────────────┐
│ Hanzo Vector Node │
├──────────────┬───────────────┬───────────────────────────┤
│ REST API │ gRPC API │ Web UI │
│ (actix-web) │ (tonic) │ (static files) │
│ :6333 │ :6334 │ :6333/dashboard │
├──────────────┴───────────────┴───────────────────────────┤
│ Storage Layer │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ Collection │ │
│ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │
│ │ │ Shard 0 │ │ Shard 1 │ │ Shard N │ │ │
│ │ │┌────────┐│ │┌────────┐│ │┌────────┐│ │ │
│ │ ││Segment ││ ││Segment ││ ││Segment ││ │ │
│ │ ││ HNSW ││ ││ HNSW ││ ││ HNSW ││ │ │
│ │ ││ Index ││ ││ Index ││ ││ Index ││ │ │
│ │ │└────────┘│ │└────────┘│ │└────────┘│ │ │
│ │ └──────────┘ └──────────┘ └──────────┘ │ │
│ └─────────────────────────────────────────────────────┘ │
├──────────────────────────────────────────────────────────┤
│ WAL (Write-Ahead Log) │ Snapshots │ Gridstore │
├──────────────────────────────────────────────────────────┤
│ Raft Consensus (P2P :6335) -- cluster mode only │
└──────────────────────────────────────────────────────────┘

Data Model

  • Collection: A named group of points with a defined vector configuration (dimension, distance metric)
  • Point: A record with an ID (u64 or UUID), one or more named vectors, and an optional JSON payload
  • Payload: Arbitrary JSON metadata attached to a point, indexable for filtering
  • Shard: A horizontal partition of a collection, distributable across cluster nodes
  • Segment: An immutable storage unit within a shard containing vectors + indexes

Distance Metrics

  • Cosine -- angular similarity (default for embeddings)
  • Euclidean -- L2 distance
  • Dot -- inner product
  • Manhattan -- L1 distance

Quantization

  • Scalar -- 8-bit quantization (4x memory reduction)
  • Product -- PQ compression (up to 64x reduction)
  • Binary -- 1-bit per dimension (32x reduction, fastest)

Workspace Structure

hanzoai/vector (master branch)
├── Cargo.toml # Root workspace, package: hanzo-vector v1.17.0
├── Cargo.lock
├── Dockerfile # Multi-stage: chef → planner → builder → runtime
├── config/
│ ├── config.yaml # Full reference config with all options
│ ├── production.yaml # Production overrides
│ └── development.yaml # Dev overrides
├── src/ # Main binary source
│ ├── main.rs # Entry point, CLI args, server startup
│ ├── consensus.rs # Raft consensus implementation
│ ├── settings.rs # Config loading and validation
│ ├── snapshots.rs # Snapshot management
│ ├── startup.rs # Server initialization
│ ├── greeting.rs # Startup banner
│ ├── actix/ # REST API handlers (actix-web)
│ │ ├── api/ # Route handlers per resource
│ │ │ ├── collections_api.rs
│ │ │ ├── search_api.rs
│ │ │ ├── query_api.rs
│ │ │ ├── recommend_api.rs
│ │ │ ├── retrieve_api.rs
│ │ │ ├── update_api.rs
│ │ │ ├── count_api.rs
│ │ │ ├── cluster_api.rs
│ │ │ ├── snapshot_api.rs
│ │ │ ├── shards_api.rs
│ │ │ ├── local_shard_api.rs
│ │ │ ├── discovery_api.rs
│ │ │ ├── facet_api.rs
│ │ │ └── service_api.rs
│ │ ├── auth.rs # API key + JWT RBAC auth
│ │ └── mod.rs # Actix server config
│ ├── tonic/ # gRPC service implementations
│ ├── tracing/ # Tracing/logging setup
│ └── migrations/ # Storage format migrations
├── lib/ # Workspace crates
│ ├── api/ # API types (REST + gRPC)
│ │ ├── src/grpc/proto/ # Protobuf definitions
│ │ │ ├── qdrant.proto
│ │ │ ├── collections.proto
│ │ │ ├── points.proto
│ │ │ ├── collections_service.proto
│ │ │ ├── points_service.proto
│ │ │ ├── snapshots_service.proto
│ │ │ ├── shard_snapshots_service.proto
│ │ │ ├── raft_service.proto
│ │ │ └── health_check.proto
│ │ └── src/rest/ # REST API types
│ ├── collection/ # Collection management (CRUD, optimization)
│ ├── segment/ # Core storage engine
│ │ └── src/
│ │ ├── index/ # HNSW index implementation
│ │ ├── vector_storage/ # Vector data storage
│ │ ├── payload_storage/# Payload storage + indexing
│ │ ├── id_tracker/ # Point ID mapping
│ │ ├── spaces/ # Distance metric implementations
│ │ ├── data_types/ # Vector type definitions
│ │ └── types.rs # Core type definitions (170KB)
│ ├── storage/ # Storage orchestration layer
│ ├── shard/ # Shard management and transfer
│ ├── sparse/ # Sparse vector support
│ ├── quantization/ # Quantization (scalar/product/binary, C++ FFI)
│ ├── posting_list/ # Inverted index for sparse vectors
│ ├── gridstore/ # Column-oriented storage for payloads
│ ├── gpu/ # GPU acceleration (Vulkan, NVIDIA/AMD)
│ ├── edge/ # Embedded mode (no server, direct API)
│ │ └── python/ # Python bindings for edge mode
│ ├── trififo/ # Lock-free triple-buffer FIFO
│ ├── macros/ # Proc macros
│ └── common/ # Shared utilities
│ ├── common/ # Core common types
│ ├── cancel/ # Cancellation tokens
│ ├── dataset/ # Dataset loading utilities
│ └── issues/ # Issue tracking/reporting
├── openapi/ # OpenAPI 3.0 schema (ytt templates)
│ ├── openapi-main.ytt.yaml
│ ├── openapi-collections.ytt.yaml
│ ├── openapi-points.ytt.yaml
│ ├── openapi-service.ytt.yaml
│ ├── openapi-snapshots.ytt.yaml
│ ├── openapi-shard-snapshots.ytt.yaml
│ ├── openapi-shards.ytt.yaml
│ ├── openapi-cluster.ytt.yaml
│ └── schemas/
├── tests/ # Integration tests
│ ├── basic_api_test.sh
│ ├── basic_grpc_test.sh
│ ├── basic_sparse_test.sh
│ ├── consensus_tests/
│ ├── e2e_tests/
│ └── openapi/
├── tools/ # Build and dev scripts
│ ├── entrypoint.sh # Docker entrypoint
│ ├── sync-web-ui.sh # Download web dashboard
│ ├── compose/ # Docker compose files for cluster
│ └── schema2openapi/ # Schema generation tooling
└── pkg/
 └── appimage/ # AppImage packaging

REST API Endpoints

All endpoints are on port 6333 by default. Auth via api-key header or JWT bearer token.

Collections

# List collections
GET /collections

# Create collection
PUT /collections/{name}
{
 "vectors": { "size": 1536, "distance": "Cosine" },
 "optimizers_config": { "indexing_threshold": 10000 },
 "replication_factor": 2
}

# Get collection info
GET /collections/{name}

# Delete collection
DELETE /collections/{name}

# Update collection params
PATCH /collections/{name}

Points

# Upsert points
PUT /collections/{name}/points
{
 "points": [
 {
 "id": 1,
 "vector": [0.1, 0.2, ...],
 "payload": { "city": "Berlin", "category": "tech" }
 }
 ]
}

# Get points by ID
POST /collections/{name}/points
{ "ids": [1, 2, 3], "with_payload": true, "with_vector": true }

# Delete points
POST /collections/{name}/points/delete
{ "points": [1, 2, 3] }

# Count points
POST /collections/{name}/points/count
{ "filter": { "must": [{ "key": "city", "match": { "value": "Berlin" } }] } }
# Vector search
POST /collections/{name}/points/search
{
 "vector": [0.1, 0.2, ...],
 "limit": 10,
 "filter": {
 "must": [{ "key": "city", "match": { "value": "Berlin" } }]
 },
 "with_payload": true
}

# Batch search
POST /collections/{name}/points/search/batch
{ "searches": [...] }

# Query (universal search endpoint)
POST /collections/{name}/points/query
{
 "query": [0.1, 0.2, ...],
 "limit": 10,
 "filter": { ... }
}

# Recommend
POST /collections/{name}/points/recommend
{
 "positive": [1, 2],
 "negative": [3],
 "limit": 10
}

# Discover (context-based search)
POST /collections/{name}/points/discover

Snapshots

# Create snapshot
POST /collections/{name}/snapshots

# List snapshots
GET /collections/{name}/snapshots

# Download snapshot
GET /collections/{name}/snapshots/{snapshot_name}

# Full storage snapshot
POST /snapshots

Cluster

# Cluster info
GET /cluster

# Collection cluster info
GET /collections/{name}/cluster

# Move shard
POST /collections/{name}/cluster
{ "move_shard": { "shard_id": 0, "from_peer_id": 1, "to_peer_id": 2 } }

Service

# Health check
GET /healthz

# Readiness
GET /readyz

# Telemetry
GET /telemetry

# Metrics (Prometheus)
GET /metrics

# Locks
GET /locks
POST /locks

gRPC Services

Port 6334 by default. Proto files in lib/api/src/grpc/proto/.

Service Proto Description
Collections collections_service.proto CRUD for collections
Points points_service.proto CRUD, search, recommend, query for points
Snapshots snapshots_service.proto Snapshot management
ShardSnapshots shard_snapshots_service.proto Per-shard snapshot operations
Qdrant qdrant.proto Health check, version info
QdrantInternal qdrant_internal_service.proto Internal cluster operations
Raft raft_service.proto Raft consensus messages

Configuration

The config file (config/config.yaml) supports YAML with environment variable overrides. Key sections:

storage:
 storage_path: ./storage # Data directory
 snapshots_path: ./snapshots # Snapshot directory
 on_disk_payload: true # Keep payloads on disk (saves RAM)
 wal:
 wal_capacity_mb: 32 # WAL segment size
 hnsw_index:
 m: 16 # HNSW edges per node
 ef_construct: 100 # Build-time neighbors
 full_scan_threshold_kb: 10000
 on_disk: false # HNSW in RAM or disk
 optimizers:
 indexing_threshold_kb: 10000
 flush_interval_sec: 5
 collection:
 replication_factor: 1
 write_consistency_factor: 1

service:
 http_port: 6333
 grpc_port: 6334
 host: 0.0.0.0
 max_request_size_mb: 32
 enable_cors: true
 enable_tls: false
 # api_key: your_secret_key
 # jwt_rbac: true

cluster:
 enabled: false
 p2p:
 port: 6335
 consensus:
 tick_period_ms: 100

telemetry_disabled: false

Environment variable override pattern: QDRANT__SERVICE__HTTP_PORT=6333 (double underscore for nesting).

Distributed Deployment

Enable cluster mode with cluster.enabled: true and specify bootstrap peers:

# Node 1 (bootstrap)
./qdrant --uri http://node1:6335

# Node 2
./qdrant --uri http://node2:6335 --bootstrap http://node1:6335

# Node 3
./qdrant --uri http://node3:6335 --bootstrap http://node1:6335

Cluster features:

  • Raft consensus for metadata coordination
  • Sharding: Automatic or manual shard distribution
  • Replication: Configurable replication factor per collection
  • Shard transfer methods: stream_records, snapshot, wal_delta
  • Write consistency: Configurable quorum (1 to all replicas)

Debug Binaries

Available with --features service_debug:

# Generate JSON schema for all types
cargo run --features service_debug --bin schema_generator

# Inspect WAL contents
cargo run --features service_debug --bin wal_inspector -- <path>

# Pop entries from WAL
cargo run --features service_debug --bin wal_pop -- <path>

# Inspect segment data
cargo run --features service_debug --bin segment_inspector -- <path>

Docker Build

The Dockerfile supports multi-platform builds with optional GPU:

# Standard build
docker build -t hanzo-vector .

# With GPU (NVIDIA)
docker build --build-arg GPU=nvidia -t hanzo-vector:gpu .

# With GPU (AMD)
docker build --build-arg GPU=amd -t hanzo-vector:gpu-amd .

# Custom profile
docker build --build-arg PROFILE=ci -t hanzo-vector:ci .

# With extra features
docker build --build-arg FEATURES=rocksdb -t hanzo-vector:rocksdb .

Feature Flags

Feature Description
gpu GPU-accelerated HNSW (NVIDIA Vulkan / AMD ROCm)
rocksdb RocksDB storage backend (alternative to default)
tracing Distributed tracing support
console tokio-console subscriber for runtime debugging
tracy Tracy profiler integration
stacktrace Stack trace capture on Linux
staging Staging environment features
data-consistency-check Runtime data integrity verification
chaos-testing Chaos engineering hooks

Testing

# Unit tests
cargo test

# Integration tests (requires running instance)
bash tests/basic_api_test.sh
bash tests/basic_grpc_test.sh
bash tests/basic_sparse_test.sh

# Consensus tests
cd tests/consensus_tests && cargo test

# E2E tests
cd tests/e2e_tests && cargo test

# OpenAPI consistency check
bash tests/openapi_consistency_check.sh

# Python integration tests (uses uv)
cd tests && uv run pytest

Troubleshooting

Issue Cause Solution
Port 6333 in use Another instance running lsof -i :6333 and kill
OOM on large collections Vectors in RAM Set on_disk: true in vector config or use quantization
Slow search No HNSW index built yet Wait for optimizer or lower indexing_threshold_kb
gRPC connection refused gRPC disabled Set grpc_port: 6334 in config
Cluster peer unreachable P2P port blocked Open port 6335, check TLS settings
GPU not detected Missing Vulkan drivers Install vulkan-tools, verify with vulkaninfo
Build fails on macOS Missing protoc brew install protobuf
  • hanzo/hanzo-database.md - PostgreSQL (pgvector for simpler vector search)
  • hanzo/hanzo-storage.md - S3-compatible object storage
  • hanzo/hanzo-llm-gateway.md - LLM Gateway (generates embeddings for vector storage)

Last Updated: 2026-03-13 Category: Hanzo Ecosystem Related: vector, embeddings, similarity-search, qdrant, hnsw, ai Prerequisites: Rust, Docker, protobuf basics