rag_conditional

Conditional RAG Example

This example demonstrates a conditional RAG pipeline with dynamic routing based on relevance scores.

Overview

The conditional RAG pattern uses intelligent routing:

1. Retrieve: Find potentially relevant documents
2. Rerank: Score documents for relevance
3. Route: Conditionally choose the next step based on relevance
   • High relevance → Generate answer directly
   • Low relevance → Trigger fallback search
4. Generate: Create answer with available context
5. Format: Add citations

This implementation is suitable for:

• Hybrid search systems
• Applications with variable query types
• Systems requiring fallback mechanisms
• Production systems needing robustness

Features

• Conditional Routing: Dynamic path selection based on relevance scores
• Relevance Threshold: Configurable threshold for routing decisions
• Fallback Search: Alternative search when primary retrieval fails
• Transparency: Shows which path was taken and why
• Adaptive Behavior: Different handling for different query types

Running the Example

cd examples/rag_conditional
go run main.go

Key Components

• Retriever: Initial document retrieval
• Reranker: Scores documents for relevance
• Conditional Edge: Routes based on relevance threshold
• Fallback Search: Alternative search mechanism
• LLM: DeepSeek-v3 for answer generation
• Pipeline: Conditional RAG with branching logic

Pipeline Flow

Query → Retrieve → Rerank → Check Relevance Score
                              ↓
                    ┌─────────┴─────────┐
                    ↓                   ↓
            Score >= Threshold    Score < Threshold
                    ↓                   ↓
              Generate           Fallback Search
                    ↓                   ↓
                    └─────────┬─────────┘
                              ↓
                      Format Citations → Result

Routing Logic

The pipeline uses a relevance threshold (default: 0.5) to decide:

• High Relevance (≥ 0.5): Documents are relevant, proceed to generation
• Low Relevance (< 0.5): Documents may not be relevant, trigger fallback

Example Queries

The example includes queries that demonstrate both paths:

1. Relevant queries: "How does checkpointing work?" → Direct generation
2. Irrelevant queries: "What is the weather?" → Fallback search

Use Cases

1. Hybrid Search: Combine vector search with keyword or web search
2. Quality Control: Only use retrieved docs if they're truly relevant
3. Graceful Degradation: Provide alternative responses for out-of-domain queries
4. Multi-Source RAG: Route to different knowledge sources based on query type

Customization

You can customize:

• Relevance threshold for routing
• Fallback search implementation (e.g., web search, keyword search)
• Routing logic and conditions
• Number of documents to retrieve
• Reranking algorithm

Best Practices

1. Tune Threshold: Adjust based on your use case and quality requirements
2. Implement Fallback: Provide meaningful fallback (web search, default response)
3. Monitor Routing: Track which path queries take for optimization
4. Test Edge Cases: Ensure good behavior for both relevant and irrelevant queries