Voyage AI: The Precision Behind Vectora

The Heart of Vectora Retrieval: Voyage 4 & Rerank 2.5

While Gemini 3 Flash provides the reasoning, the precision of Vectora’s search relies on Voyage AI. Specifically, we use Voyage 4 for embeddings and Voyage Rerank 2.5 to ensure that only the most relevant code snippets reach the LLM.

Why Voyage AI?

We tested leading embedding models. Here is the reality for code retrieval:

Embedding Model Comparison (Code Retrieval)

Why Voyage Wins for Developers

1. Code-Specific Tuning: Voyage 4 is trained specifically on vast repositories of source code (Python, TS, Go, Rust), understanding complex syntax better than general-purpose models.
2. Efficiency: 1,536 dimensions provide the perfect balance between retrieval precision and storage cost.
3. Large Window: 32K token support allows for much larger “chunks” or files to be processed without losing context.
4. Reranking Integration: Voyage Rerank 2.5 works seamlessly with Voyage 4, reducing noise and increasing the “precision at 1” (P@1).

Architecture: The Vector Pipeline

Vectora uses a two-stage retrieval pipeline:

1. Vector Search (Coarse Retrieval)

1. Chunking: Your code is split into optimized chunks with Tree-sitter awareness.
2. Embedding: Voyage 4 converts these chunks into high-dimensional vectors.
3. Indexing: Vectors are stored in a Qdrant HNSW index.
4. Query: When you search, your question is embedded, and Qdrant finds the top 50-100 nearest neighbors.

2. Reranking (Fine Filtering)

Vector search alone can return chunks that are semantically similar but technically irrelevant. Voyage Rerank 2.5 re-evaluates the top results:

• Semantic Match: Does this code actually solve the user’s problem?
• Contextual Relevance: Is this the right file or just a similar syntax pattern?
• Result: The top 5 highest-scoring chunks are sent to Gemini.

Benefits of Reranking

Without a reranker, the LLM often gets “confused” by similar-looking but incorrect code. Reranking provides:

• Higher Quality: Reduces hallucination by providing only true positives.
• Lower Cost: Reduces the tokens sent to the LLM by filtering out fluff.
• Better UX: Answers feel more “aware” of your specific project structure.

Configuration

To use Voyage AI, ensure your vectora.config.yaml is configured:

providers:
  embedding:
    name: "voyage"
    model: "voyage-4"
    api_key: "${VOYAGE_API_KEY}"

  reranker:
    name: "voyage"
    model: "voyage-rerank-2.5"
    api_key: "${VOYAGE_API_KEY}"

Pricing & Quotas

Voyage AI is extremely cost-effective:

Most small to medium projects cost less than $1/month in Voyage API fees on the Free plan (BYOK).

Part of the Vectora ecosystem · Open Source (MIT) · Contributors