Reranker Module

The Reranker module provides concrete implementations for reranking search results to improve relevance.

Import

from praisonai.adapters import LLMReranker
from praisonai.adapters.rerankers import CrossEncoderReranker, CohereReranker

Quick Example

For reranker registry lookups, call register_default_adapters() once at startup — importing reranker classes no longer auto-registers them (PraisonAI 4.6.156+). See Registering the built-in adapters.

from praisonai.adapters import LLMReranker

# Create LLM-based reranker
reranker = LLMReranker(model="gpt-4o-mini")

# Rerank documents
documents = ["Python is a language", "Java is compiled", "Python uses indentation"]
results = reranker.rerank(
    query="What is Python?",
    documents=documents,
    top_k=2
)

for r in results:
    print(f"Score: {r.score:.3f} - {r.text}")

Features

LLM-based relevance scoring with any model
Cross-encoder neural reranking
Cohere Rerank API integration
Batch processing for efficiency
Async support

Classes

`LLMReranker`

Uses an LLM to score document relevance to a query.

from praisonai.adapters import LLMReranker

reranker = LLMReranker(
    model="gpt-4o-mini",
    batch_size=5
)
results = reranker.rerank("query", documents)

Parameters:

Parameter	Type	Default	Description
`llm`	`Any`	`None`	Custom LLM instance (Agent)
`model`	`str`	`"gpt-4o-mini"`	Model for scoring
`batch_size`	`int`	`5`	Documents per batch

How it works:

Prompts the LLM to rate relevance on a 0-10 scale
Normalizes scores to 0-1 range
Sorts by score descending

`CrossEncoderReranker`

Uses sentence-transformers cross-encoder for accurate relevance scoring.

from praisonai.adapters.rerankers import CrossEncoderReranker

reranker = CrossEncoderReranker(
    model_name="cross-encoder/ms-marco-MiniLM-L-6-v2"
)
results = reranker.rerank("query", documents, top_k=5)

Parameters:

Parameter	Type	Default	Description
`model_name`	`str`	`"cross-encoder/ms-marco-MiniLM-L-6-v2"`	Cross-encoder model

Requires sentence-transformers package: pip install sentence-transformers

`CohereReranker`

Uses Cohere’s rerank API for high-quality reranking.

from praisonai.adapters.rerankers import CohereReranker

reranker = CohereReranker(
    api_key="your-api-key",  # Or set COHERE_API_KEY env
    model="rerank-english-v3.0"
)
results = reranker.rerank("query", documents, top_k=5)

Parameters:

Parameter	Type	Default	Description
`api_key`	`str`	`COHERE_API_KEY` env	Cohere API key
`model`	`str`	`"rerank-english-v3.0"`	Rerank model

Requires cohere package: pip install cohere

Methods

`rerank(query, documents, top_k=None)`

Rerank documents by relevance to query. Parameters:

query (str): Search query
documents (List[str]): Documents to rerank
top_k (int, optional): Number of results to return

Returns: List[RerankResult] - Reranked documents with scores

`arerank(query, documents, top_k=None)`

Async reranking that offloads the blocking body via asyncio.to_thread(...), so calling it from an event loop no longer freezes it. Authors of custom rerankers on a natively-async provider should override arerank and await the native call directly — see Native async in custom retrievers / rerankers.

Example: Full RAG Pipeline

from praisonai.adapters import (
    ChromaVectorStore, 
    BasicRetriever, 
    LLMReranker
)

# Setup
store = ChromaVectorStore(namespace="docs")
retriever = BasicRetriever(
    vector_store=store,
    embedding_fn=get_embedding,
    top_k=20  # Retrieve more for reranking
)
reranker = LLMReranker(model="gpt-4o-mini")

# Retrieve
query = "How to deploy Python apps?"
results = retriever.retrieve(query)

# Rerank
documents = [r.text for r in results]
reranked = reranker.rerank(query, documents, top_k=5)

# Use top results
for r in reranked:
    print(f"Score: {r.score:.3f}")
    print(f"Text: {r.text[:100]}...")

Example: Using with Agent

from praisonai.adapters import LLMReranker
from praisonaiagents import Agent

# Use an agent for reranking
agent = Agent(
    instructions="You are a relevance scoring assistant",
    llm="gpt-4o-mini"
)

reranker = LLMReranker(llm=agent)
results = reranker.rerank("Python basics", documents)

Reranker Selection Guide

Use Case	Recommended
Best accuracy	`CohereReranker`
Local/offline	`CrossEncoderReranker`
Already using LLM	`LLMReranker`
Fast, no dependencies	Skip reranking

Environment Variables

# For LLM reranker (uses litellm)
export OPENAI_API_KEY=sk-xxx

# For Cohere reranker
export COHERE_API_KEY=xxx

CLI Usage

# LLM reranker
praisonai knowledge query "What is Python?" --reranker llm

# Cross-encoder reranker
praisonai knowledge query "What is Python?" --reranker cross_encoder

# Cohere reranker
praisonai knowledge query "What is Python?" --reranker cohere

Retrieval Module - Retrieve documents
Vector Store Module - Store documents

​Reranker Module

​Import

​Quick Example

​Features

​Classes

​LLMReranker

​CrossEncoderReranker

​CohereReranker

​Methods

​rerank(query, documents, top_k=None)

​arerank(query, documents, top_k=None)

​Example: Full RAG Pipeline

​Example: Using with Agent

​Reranker Selection Guide

​Environment Variables

​CLI Usage

​Related

Reranker Module

Import

Quick Example

Features

Classes

`LLMReranker`

`CrossEncoderReranker`

`CohereReranker`

Methods

`rerank(query, documents, top_k=None)`

`arerank(query, documents, top_k=None)`

Example: Full RAG Pipeline

Example: Using with Agent

Reranker Selection Guide

Environment Variables

CLI Usage

Related