coreason-search

The Unified Retrieval Execution Engine for the CoReason ecosystem.

coreason-search serves as the "Librarian" of the CoReason architecture, designed to solve three distinct problems: Ad-Hoc Reasoning (RAG), Systematic Evidence Synthesis, and Context Distillation. It adopts a "Late Chunking / No Chunking" philosophy, utilizing SOTA 32k context embeddings to process full documents.

Features

• Long-Context Sovereignty: Utilizes SOTA embeddings (e.g., Qwen3-Embedding) with 32k context windows to ingest full research papers without "chunk-and-pray" methods.
• Modular Strategy Pattern: Supports multiple retrieval strategies:
  • Dense Retriever: Vector-based semantic search using LanceDB.
  • Sparse Retriever: Boolean/Keyword search for systematic reviews (PubMed style).
  • Graph Retriever: Symbolic 1-hop neighbor expansion using coreason-graph-nexus.
• Hybrid Fusion: Implements Reciprocal Rank Fusion (RRF) to merge results from vector and keyword searches robustly.
• Precision Re-Ranking: Uses Cross-Encoders to re-rank top results for maximum precision.
• Context Distillation (The Scout): Compresses documents by stripping irrelevant sentences, maximizing the signal-to-noise ratio for downstream LLMs.
• Systematic Search Mode: Supports "Research-Grade" reproducible reviews with strict boolean logic and generator-based pagination for large result sets.

Installation

pip install coreason-search

Usage

from coreason_search.engine import SearchEngine
from coreason_search.schemas import SearchRequest, RetrieverType

# Initialize the engine (loads configuration automatically)
engine = SearchEngine()

# Define a search request
request = SearchRequest(
    query="mechanism of action of aspirin in liver failure",
    strategies=[RetrieverType.LANCE_DENSE, RetrieverType.LANCE_FTS],
    fusion_enabled=True,
    rerank_enabled=True,
    distill_enabled=True,
    top_k=5
)

# Execute the search
response = engine.search(request)

# Process results
for hit in response.hits:
    print(f"[{hit.score:.4f}] {hit.doc_id}")
    print(f"Distilled Context: {hit.distilled_text}\n")