Skip to content

❓ FAQ

Quick answers to the most common questions about Spector. Can't find what you're looking for? Check GitHub Discussions or the specific wiki pages linked throughout.

🌟 General

What Java version do I need?

JDK 25 or later. Spector uses the Java Vector API (incubator module) for SIMD acceleration and Panama FFM for off-heap memory. OpenJDK builds include these by default.

Does it work without a GPU?

Yes, completely. GPU is optional. Without a GPU, Spector uses CPU SIMD acceleration (AVX2/AVX-512/NEON) which delivers sub-millisecond search at 100K documents. GPU helps primarily for high-concurrency batch workloads.

Tip

See GPU Acceleration for details on when GPU adds value (spoiler: batch sizes > 32).

Can I use it as an embedded library?

Absolutely! Spector runs in two modes:

Mode Description Overhead
Embedded Add JAR to classpath, create SpectorEngine Zero network overhead
Server REST API with auth, CORS, metrics HTTP overhead 
try (var engine = new SpectorEngine(SpectorConfig.DEFAULT.withDimensions(384))) {
    engine.ingest("id", "content", vector);
    var results = engine.hybridSearch("query", queryVector, 10);
}

What about persistence? Do I lose data on restart?

No! Spector supports persistence through memory-mapped files. The HNSW index uses a page-aligned binary format that loads instantly via mmap — no deserialization needed. Vector data survives restarts.

How does it compare to Elasticsearch?

Aspect ⚡ Spector Elasticsearch
Vector search latency 0.13 ms (100K, in-process) 2–10 ms
Hybrid search latency 1.01 ms (100K, in-process) 10–30 ms
Deployment Embedded JAR or server Cluster only
Dependencies Zero (JDK only) JVM + heavy stack
GPU support ✅ CUDA ❌
IVF-PQ compression ✅ 32× ❌

Elasticsearch excels at distributed full-text search with a mature query language and ecosystem. Spector excels at raw in-process performance, embedded use, and modern JVM features. The latency advantage is largest for in-process embedded use; network-bound deployments narrow the gap.

Does it support filtering/metadata queries?

Yes. The Spring AI integration supports filter expressions:

vectorStore.similaritySearch(
    SearchRequest.query("search algorithms")
        .withFilterExpression("category == 'indexing' && version > 2")
);

What embedding models work with Spector?

Any model that produces float32 vectors. Set dimensions to match:

Model Dimensions Provider
all-MiniLM-L6-v2 384 Sentence Transformers / Ollama
e5-base-v2 768 Sentence Transformers
text-embedding-ada-002 1536 OpenAI
nomic-embed-text 768 Ollama
mxbai-embed-large 1024 Ollama

Note

Spector includes an Ollama embedding provider out of the box. Implement the EmbeddingProvider SPI for any other source.

🔧 Technical

What similarity functions are supported?

Function Best For
COSINE (default) Normalized embeddings (most models)
DOT_PRODUCT Unnormalized embeddings, magnitude matters
EUCLIDEAN Spatial/geometric data

What's the maximum dataset size?

Mode Scale
Single node Up to 10 million documents
IVF-PQ mode Billions of vectors (32× compression)
Distributed mode Scale horizontally (2–256 shards)

How does the LLM re-ranking work?

flowchart LR
    A["🔍 Search<br/>Top-N candidates"] --> B["🤖 LLM (Ollama)<br/>Listwise scoring"]
    B --> C["✨ Re-ranked<br/>Top-K results"]
  1. Vector/hybrid search retrieves top-N candidates (default: 20)
  2. Candidates sent to Ollama for listwise relevance scoring
  3. LLM reorders based on semantic relevance
  4. Final top-K results reflect LLM judgment

Warning

Adds 100–500ms latency but significantly improves precision for ambiguous queries.

What are virtual threads and why do they matter?

Virtual threads (Project Loom) are lightweight threads that don't map 1:1 to OS threads:

  • ✅ Handle millions of concurrent requests without pool tuning

  • ✅ No synchronized blocks that pin platform threads

  • ✅ Near-zero scheduling overhead

  • ✅ Linear scaling (4.5× at 16 threads measured)

How does zero-copy storage work?

Vectors are stored in memory-mapped files using Panama's MemorySegment:

  • OS maps file directly into process address space

  • SIMD kernels read vectors without copying to Java heap

  • Zero garbage collection pressure

  • Instant startup (no deserialization)

  • Supports datasets larger than available RAM

What's the difference between HNSW and IVF-PQ?

Aspect 🌐 HNSW 🗜️ IVF-PQ
Speed Fastest (0.05ms) Fast (nprobe-dependent)
Memory Full vectors (1.5KB/vec @ 384-dim) 32× compressed (48 bytes/vec)
Recall High (configurable) Moderate (nprobe-dependent)
Scale Up to millions Up to billions
Use case Default for most workloads Memory-constrained, billion-scale

Can I run benchmarks in CI?

Yes! JSON output + baseline regression detection:

mvn -pl spector-bench exec:java -Dexec.args="-rf json -rff results.json"

⚙️ Operations

What ports does Spector use?

Port Protocol Purpose
7070 HTTP REST API (configurable)
9090 gRPC Cluster communication (distributed mode)

How do I monitor Spector?

curl http://localhost:7070/health          # Health check
curl http://localhost:7070/api/v1/status    # Engine status
curl http://localhost:7070/api/v1/metrics   # Request metrics

What JVM arguments should I use in production?

java \
  --add-modules jdk.incubator.vector \
  --enable-native-access=ALL-UNNAMED \
  -XX:+UseZGC -XX:+ZGenerational \
  -Xmx4g -Xms4g \
  -jar spector-node.jar

How do I upgrade without downtime?

Distributed mode: 1. Drain one node (stop routing requests) 2. Upgrade the node binary 3. Restart and wait for replica sync 4. Repeat for each node

Embedded mode: Standard application deployment with new Spector version.

Is there authentication?

Yes. Set an API key at server startup:

mvn exec:java -pl spector-node \
  -Dexec.args="7070 384 my-secret-key"

Clients include X-API-Key: my-secret-key in requests. Without a key configured, all requests are allowed.

🔗 See Also