query-decomposition

# Query Decomposition
Break complex queries into independent concepts for parallel retrieval and fusion.

## Overview

- Complex queries spanning multiple topics or concepts
- Multi-hop questions requiring chained reasoning
- Queries where single retrieval misses relevant documents
- Improving recall for compound questions

Break complex queries into independent concepts for parallel retrieval and fusion.

## The Problem

Complex queries span multiple topics that may not co-occur in single documents:
```
Query: "How do chunking strategies affect reranking in RAG?"
→ Single search may miss docs about chunking OR reranking
→ Poor coverage across all concepts
```

## The Solution

Decompose into independent concepts, retrieve separately, then fuse:
```
Query: "How do chunking strategies affect reranking in RAG?"
→ Concepts: ["chunking strategies", "reranking methods", "RAG pipeline"]
→ Search each concept independently
→ Fuse results with Reciprocal Rank Fusion (RRF)
→ Full coverage across all topics
```

## Implementation

### 1. Heuristic Detection (Fast Path)

```python
MULTI_CONCEPT_INDICATORS = [
    " vs ", " versus ", " compared to ", " or ",
    " and ", " with ", " affect ", " impact ",
    "difference between", "relationship between",
]

def is_multi_concept_heuristic(query: str) -> bool:
    """Fast check for multi-concept indicators (<1ms)."""
    query_lower = query.lower()
    return any(ind in query_lower for ind in MULTI_CONCEPT_INDICATORS)
```

### 2. LLM Decomposition

```python
from pydantic import BaseModel, Field
from openai import AsyncOpenAI

class ConceptExtraction(BaseModel):
    """LLM output schema for concept extraction."""
    concepts: list[str] = Field(
        ...,
        min_length=1,
        max_length=5,
        description="Distinct concepts from the query",
    )
    reasoning: str | None = None

async def decompose_query(
    query: str,
    llm: AsyncOpenAI,
) -> list[str]:
    """Extract independent concepts using LLM."""

    r