Implementation:Hpcaitech ColossalAI RetrievalConversation

Knowledge Sources	ColossalAI
Domains	RAG, NLP
Last Updated	2026-02-09 00:00 GMT

Overview

Conversational RAG pipeline with query disambiguation and rejection filtering, provided by ColossalQA.

Description

ChineseRetrievalConversation and EnglishRetrievalConversation implement language-specific RAG pipelines. They combine a CustomRetriever, an LLM (via ColossalAPI), conversation memory (ConversationBufferWithSummary), a disambiguation chain, and a RetrievalQA chain into a complete conversational QA system.

Usage

Create from a retriever and LLM configuration, then call run() with user queries.

Code Reference

Source Location

• Repository: ColossalAI
• File (Chinese): applications/ColossalQA/colossalqa/retrieval_conversation_zh.py
• Lines: 23-96
• File (English): applications/ColossalQA/colossalqa/retrieval_conversation_en.py
• Lines: 23-88

Signature

class ChineseRetrievalConversation:
    def __init__(
        self,
        retriever: CustomRetriever,
        model_path: str,
        model_name: str,
    ) -> None:
        """
        Args:
            retriever: Configured CustomRetriever with indexed documents
            model_path: Path to local LLM model
            model_name: Model type identifier (e.g., "chatglm2")
        """

    def run(
        self,
        user_input: str,
        memory: ConversationBufferWithSummary,
    ) -> Tuple[str, ConversationBufferWithSummary]:
        """Process a user query through the RAG pipeline."""

    @classmethod
    def from_retriever(cls, retriever, model_path, model_name):
        """Factory method to create from a retriever."""

Import

from colossalqa.retrieval_conversation_zh import ChineseRetrievalConversation
from colossalqa.retrieval_conversation_en import EnglishRetrievalConversation

I/O Contract

Inputs

Name	Type	Required	Description
retriever	CustomRetriever	Yes	Configured retriever with indexed documents
model_path	str	Yes	Path to local LLM
model_name	str	Yes	Model type identifier
user_input	str	Yes	User's question
memory	ConversationBufferWithSummary	Yes	Conversation history

Outputs

Name	Type	Description
answer	str	Generated answer (first line of response)
memory	ConversationBufferWithSummary	Updated conversation memory

Usage Examples

from colossalqa.retrieval_conversation_zh import ChineseRetrievalConversation
from colossalqa.memory import ConversationBufferWithSummary

conversation = ChineseRetrievalConversation(
    retriever=retriever,
    model_path="/models/chatglm2-6b",
    model_name="chatglm2",
)

memory = ConversationBufferWithSummary(llm=conversation.llm)
answer, memory = conversation.run("什么是ColossalAI?", memory)
print(answer)

Related Pages

Implements Principle

• Principle:Hpcaitech_ColossalAI_Retrieval_Augmented_QA

Environment and Heuristic Links

• Environment:Hpcaitech_ColossalAI_ColossalQA_RAG_Environment