AgentSkillsCN

Redis Vectorstore

Redis Vectorstore

SKILL.md

Skill: Redis Vector Store

When to Use This Skill

Use this skill when:

  • •Implementing the knowledge base for PatchPilot
  • •Storing failure traces with embeddings
  • •Retrieving similar past failures
  • •Storing and retrieving fix patterns
  • •Integrating Redis with Triage and Fixer agents

Do NOT use this skill when:

  • •Working on browser automation (use browserbase-stagehand)
  • •Implementing deployment logic (use vercel-deployment)

Overview

Redis Stack provides vector search capabilities that enable semantic similarity search over failure traces and fix patterns. This is the "memory" that makes PatchPilot self-improving - it learns from past bugs to fix similar issues faster.

Key capabilities:

  • •Store failure traces with vector embeddings
  • •Semantic search for similar past failures
  • •Associate fixes with failure patterns
  • •Fast retrieval for real-time agent use

Key Concepts

Vector Embeddings

We convert error messages and stack traces into numerical vectors (embeddings) using OpenAI's embedding model. Similar errors will have similar vectors, enabling semantic search.

typescript
const embedding = await openai.embeddings.create({
  model: 'text-embedding-3-small',
  input: errorMessage
});
// Returns 1536-dimensional vector

Redis Vector Index

Redis Stack supports HNSW (Hierarchical Navigable Small World) indexes for efficient similarity search.

typescript
await redis.call(
  'FT.CREATE', 'failure_idx',
  'ON', 'HASH',
  'PREFIX', '1', 'failure:',
  'SCHEMA',
    'embedding', 'VECTOR', 'HNSW', '6',
      'TYPE', 'FLOAT32',
      'DIM', '1536',
      'DISTANCE_METRIC', 'COSINE',
    'error_message', 'TEXT',
    'failure_type', 'TAG',
    'file', 'TAG'
);

KNN Search

Find the K nearest neighbors to a query vector:

typescript
const results = await redis.call(
  'FT.SEARCH', 'failure_idx',
  '*=>[KNN 5 @embedding $vec AS score]',
  'PARAMS', '2', 'vec', queryVector,
  'SORTBY', 'score',
  'RETURN', '3', 'error_message', 'fix_diff', 'score'
);

Common Patterns

Store Failure Trace

typescript
async function storeFailure(
  failure: FailureTrace,
  embedding: number[]
): Promise<void> {
  const id = `failure:${failure.id}`;
  const vectorBuffer = Buffer.from(new Float32Array(embedding).buffer);

  await redis.hSet(id, {
    embedding: vectorBuffer,
    error_message: failure.errorMessage,
    stack_trace: failure.stackTrace,
    file: failure.file,
    line: failure.line.toString(),
    failure_type: failure.failureType,
    fix_description: failure.fix?.description || '',
    fix_diff: failure.fix?.diff || '',
    success: failure.fix?.success ? '1' : '0',
    created_at: Date.now().toString()
  });
}

Search Similar Failures

typescript
async function findSimilarFailures(
  errorMessage: string,
  k: number = 5
): Promise<SimilarFailure[]> {
  // Generate embedding for query
  const embedding = await generateEmbedding(errorMessage);
  const vectorBuffer = Buffer.from(new Float32Array(embedding).buffer);

  // KNN search
  const results = await redis.call(
    'FT.SEARCH', 'failure_idx',
    '*=>[KNN $k @embedding $vec AS score]',
    'PARAMS', '4',
      'k', k.toString(),
      'vec', vectorBuffer,
    'SORTBY', 'score',
    'RETURN', '5', 'error_message', 'fix_description', 'fix_diff', 'file', 'score',
    'DIALECT', '2'
  );

  return parseSearchResults(results);
}

Update with Fix

typescript
async function recordSuccessfulFix(
  failureId: string,
  fix: { description: string; diff: string }
): Promise<void> {
  const id = `failure:${failureId}`;

  await redis.hSet(id, {
    fix_description: fix.description,
    fix_diff: fix.diff,
    success: '1'
  });
}

Best Practices

  1. •Use consistent embedding model - Always use the same model (text-embedding-3-small)
  2. •Include relevant context in embeddings - Combine error message + stack trace
  3. •Filter by failure type - Use TAG filters to narrow search
  4. •Set similarity threshold - Only use matches with score > 0.8
  5. •Clean old entries - Set TTL or periodically prune

Common Pitfalls

Embedding Dimension Mismatch

  • •Index expects 1536 dimensions
  • •Ensure embedding model outputs correct size
  • •text-embedding-3-small = 1536 dim

Buffer Encoding

  • •Redis expects Float32 buffer
  • •Use Buffer.from(new Float32Array(embedding).buffer)
  • •Don't send raw JavaScript array

Index Not Created

  • •Create index before first search
  • •Use FT.INFO to check if index exists
  • •Handle "Index not found" error gracefully

Related Skills

  • •patchpilot-agents/ - Using Redis in Triage and Fixer agents
  • •wandb-weave/ - Logging Redis query results

References