Context Retrieval

Context Retrieval

Retrieve relevant episodic context from memory for informed decision-making.

Purpose

Find and return the most relevant past episodes and patterns to inform current task execution.

Retrieval Methods

1. Semantic Search (Preferred)

When embedding service is configured, use vector similarity:

let context = memory
    .retrieve_relevant_context(
        "implement async batch updates",
        task_context,
        limit: 5
    )
    .await?;

Advantages:

• Finds semantically similar tasks
• Better recall for novel queries
• Captures intent beyond keywords

Requirements:

• Embedding service configured
• Embeddings computed for episodes
• Vector index built

2. Keyword/Index Search (Fallback)

When embeddings unavailable, use SQL indexes:

// Search by task_type and tags
SELECT * FROM episodes
WHERE task_type = ?
  AND tags LIKE '%' || ? || '%'
ORDER BY timestamp DESC
LIMIT ?;

Advantages:

• Fast, no embedding computation
• Works without external services
• Deterministic results

Limitations:

• Misses semantic similarity
• Requires exact tag matches
• Vocabulary mismatch issues

Retrieval Strategy

1. Parse query: Extract key terms, domain, task type
2. Check embedding availability: Prefer semantic if available
3. Query storage:
   • Try redb cache first (fast)
   • Fall back to Turso if cache miss
4. Rank results: By relevance score or recency
5. Filter and limit: Apply filters, return top-N
6. Format context: Structure for consumption

Context Filtering

By Domain

let context = TaskContext {
    domain: "storage".to_string(),
    ..Default::default()
};

Returns only episodes in the "storage" domain.

By Task Type

// Find similar implementation tasks
task_type_filter: Some("implementation")

By Recency

// Last 30 days only
since: Some(now - Duration::days(30))

By Success

// Only successful episodes
verdict: Some(Verdict::Success)

Result Ranking

Semantic (Embedding-based)

1. Compute query embedding
2. Calculate cosine similarity with stored embeddings
3. Sort by similarity score (descending)
4. Return top-N

Index-based

1. Exact match on task_type (highest priority)
2. Tag overlap count
3. Recency (timestamp DESC)
4. Success rate

Response Format

pub struct RetrievedContext {
    pub episodes: Vec<EpisodeSummary>,
    pub patterns: Vec<Pattern>,
    pub heuristics: Vec<Heuristic>,
    pub relevance_scores: Vec<f32>,
}

pub struct EpisodeSummary {
    pub id: String,
    pub task_description: String,
    pub verdict: Verdict,
    pub key_steps: Vec<String>,
    pub reflection: String,
    pub artifacts: Vec<String>,
    pub relevance: f32,
}

Usage Examples

Retrieve for Similar Task

let query = "fix failing async test with Tokio runtime";
let context = TaskContext {
    language: "rust".to_string(),
    domain: "testing".to_string(),
    tags: vec!["tokio".to_string(), "async".to_string()],
};

let retrieved = memory
    .retrieve_relevant_context(query, context, 10)
    .await?;

// Use retrieved context to inform approach
for episode in retrieved.episodes {
    println!("Similar task: {}", episode.task_description);
    println!("Outcome: {:?}", episode.verdict);
    println!("Key insight: {}", episode.reflection);
}

Retrieve Patterns for Code Generation

// Find common tool sequences for storage implementation
let patterns = memory
    .get_patterns_by_type("ToolSequence")
    .filter(|p| p.domain == "storage")
    .await?;

// Apply learned patterns
for pattern in patterns {
    if pattern.success_rate > 0.8 {
        println!("Recommended approach: {:?}", pattern.sequence);
    }
}

Retrieve Error Solutions

// Find how similar errors were resolved
let query = "error: cannot borrow as mutable while immutable borrow exists";
let solutions = memory
    .retrieve_error_resolutions(query, 5)
    .await?;

for solution in solutions {
    println!("Resolution: {}", solution.fix);
    println!("Success rate: {}", solution.success_rate);
}

Performance Optimization

Cache Frequently Queried Context

• Keep hot patterns in memory
• Cache common query results
• Use LRU eviction policy

Limit Result Size

• Return summaries, not full episodes
• Paginate large result sets
• Stream results when possible

Index Optimization

• Compound indexes on (task_type, timestamp)
• Tag GIN/JSONB indexes for flexible filtering
• Regularly update statistics

Embedding Service Configuration

[embedding]
enabled = true
provider = "openai"  # or "local", "huggingface"
model = "text-embedding-3-small"
cache_embeddings = true
batch_size = 32

Troubleshooting

Low recall:

• Check embeddings are computed
• Verify task_type is set correctly
• Expand tag vocabulary
• Increase limit parameter

Slow retrieval:

• Check redb cache is populated
• Verify indexes exist and are used
• Monitor Turso latency
• Consider smaller result sets

Poor relevance:

• Use semantic search if available
• Improve query descriptions
• Add more context tags
• Filter by domain/task type

Monitoring Metrics

• Retrieval latency (p50, p95, p99)
• Cache hit rate
• Embedding computation time
• Result relevance scores
• Query patterns (for optimization)