Context Engine

The Context Engine is the heart of Vectora’s orchestration. It decides what, how, and when to fetch context from your codebase, avoiding noise and overfetch.

Context Engine is not just search. It is an intelligent pipeline: Embed → Search → Rerank → Compose → Validate.

The Problem

Generic agents return 50 irrelevant files for a simple query. The Context Engine filters by relevance, reducing this to 5-10 highly useful chunks.

Search Strategies

The Context Engine offers three search strategies, either independent or combined, depending on the query type and desired precision.

Semantic

Uses embeddings to find functional similarity. Ideal for queries like “How to validate tokens?”

Structural

Uses AST parsing for code relationships. Ideal for “Which functions call X?”

Hybrid

Combines semantic + structural search. Ideal for module refactoring.

Pipeline

Embedding: Query → 1024D vector (Voyage 4)
Search: Qdrant with filters by namespace
Reranking: Voyage Rerank 2.5 refines top-50 to top-10
Compaction: head/tail reduction without losing context
Validation: Guardian + Harness validate the output

Configuration

context_engine:
  strategy: "auto"
  max_depth: 3
  compaction: true
  include_ast: true
  include_dependencies: true

Practical Examples

Below are two detailed examples showing how the Context Engine processes queries and returns structured context.

Example 1: Semantic Search

Query: “How to validate tokens?”

Input:
  - Query: "How to validate tokens?"
  - Strategy: semantic
  - Namespace: your-project
  - Top-k: 10

Processing:
  1. Embed: Query → 1536D vector via Voyage 4
  2. Search: HNSW searches for 100 closest candidates
  3. Rerank: Voyage Rerank 2.5 refines to top-10
  4. Compaction: Reduces size from 15KB → 4KB while maintaining context
  5. Validate: Harness validates output, captures metrics

Output:
  chunks: [
    {file: "src/auth/jwt.ts", precision: 0.89, content: "...validateToken..."},
    {file: "src/auth/guards.ts", precision: 0.78, content: "...middleware..."},
    ...
  ]
  metadata: {
    retrieval_precision: 0.87,
    latency_ms: 234,
    total_searched: 3159,
    compaction_ratio: 0.27
  }

Example 2: Structural Search

Query: “Who calls getUserById?”

Input:
  - Symbol: getUserById
  - Strategy: structural
  - Include indirect: true

Processing:
  1. AST Parse: Analyzes file where getUserById is defined
  2. Call Graph: Finds all references (direct + indirect)
  3. Context: Extracts context lines for each call

Output:
  direct_calls: 47
  indirect_calls: 12
  callers: [
    {file: "src/middleware/auth.ts", line: 34, type: "direct"},
    {file: "src/routes/profile.ts", line: 12, type: "indirect via getUserData"},
    ...
  ]

Next: Harness Runtime

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

Local Reranker Connected RAG: Why isolated context kills your productivity