Implementation:FlagOpen FlagEmbedding Search Demo Tool

Knowledge Sources	FlagOpen_FlagEmbedding
Domains	Search_System, Conversational_AI, Hybrid_Retrieval
Last Updated	2026-02-09 00:00 GMT

Overview

⚠️ LEGACY CODE: This file is in `research/old-examples/` and is superseded by the main FlagEmbedding inference APIs.

Interactive conversational search agent combining BM25, dense retrieval, and GPT-based query rewriting and answer generation.

Description

This module implements a complete conversational search system with multiple components:

LocalDatasetLoader manages corpus data:

• Loads Wikipedia documents (title + text)
• Stores pre-computed dense embeddings for FAISS search

QueryGenerator rewrites conversational queries:

• Uses GPT-3.5 to rewrite context-dependent queries into standalone questions
• Replaces pronouns and implicit references with explicit entities
• Maintains conversation history for context

AnswerGenerator produces final answers:

• Takes rewritten query and retrieved passages
• Generates comprehensive answers using GPT-3.5
• Formats responses based on provided references

BMVectorIndex implements hybrid retrieval:

• BM25 first-stage retrieval (top 1000 candidates via Pyserini)
• Dense reranking using BGE embeddings (top N from candidates)
• Combines keyword matching with semantic similarity
• Supports Chinese with optional instruction prefixes

Agent orchestrates the full pipeline:

• Maintains conversation memory with sliding window (3500 tokens via tiktoken)
• Generates queries → retrieves passages → generates answers
• Updates memory with Q&A pairs for multi-turn coherence

The system provides verbose mode for debugging, showing rewritten queries and retrieved passages.

Usage

Use this for building interactive search demos with conversational query understanding and context-aware answer generation over large document collections.

Code Reference

Source Location

• Repository: FlagOpen_FlagEmbedding
• File: research/old-examples/search_demo/tool.py
• Lines: 1-130

Signature

class LocalDatasetLoader:
    def __init__(self, data_path, embedding_path)

class QueryGenerator:
    def __init__(self)
    def run(self, history, question)

class AnswerGenerator:
    def __init__(self)
    def run(self, history, question, references)

class BMVectorIndex:
    def __init__(self, model_path, bm_index_path, data_loader)
    def search_for_doc(self, query: str, RANKING: int = 1000, TOP_N: int = 5)

class Agent:
    def __init__(self, index)
    def answer(self, question, RANKING=1000, TOP_N=5, verbose=True)

Import

from research.old_examples.search_demo.tool import Agent, BMVectorIndex, LocalDatasetLoader

I/O Contract

Inputs

Name	Type	Required	Description
data_path	str	Yes	Path to Wikipedia dataset directory
embedding_path	str	Yes	Path to pre-computed embeddings .npy file
model_path	str	Yes	Path to BGE embedding model
bm_index_path	str	Yes	Path to Lucene BM25 index
question	str	Yes	User's conversational question
history	str	No	Previous Q&A pairs in conversation
RANKING	int	No	BM25 candidates to retrieve (default: 1000)
TOP_N	int	No	Final passages to use (default: 5)

Outputs

Name	Type	Description
answer	str	Generated answer to the question
rewritten_query	str	Standalone query (if verbose=True)
references	str	Retrieved passages (if verbose=True)

Usage Examples

from research.old_examples.search_demo.tool import Agent, BMVectorIndex, LocalDatasetLoader

# Initialize data loader
data_loader = LocalDatasetLoader(
    data_path="./search_demo_data/dataset",
    embedding_path="./search_demo_data/emb/data.npy"
)

# Initialize hybrid search index
search_index = BMVectorIndex(
    model_path="BAAI/bge-base-zh-v1.5",
    bm_index_path="./search_demo_data/index",
    data_loader=data_loader
)

# Create conversational agent
agent = Agent(index=search_index)

# Single-turn Q&A
agent.answer("什么是人工智能？", TOP_N=3, verbose=True)
# 输出: 查询：什么是人工智能？
#      知识：[passages]
#      答：人工智能是...

# Multi-turn conversation (with context)
agent.answer("它有哪些应用？", TOP_N=3, verbose=True)
# Automatically uses conversation history to understand "它" refers to AI
# 输出: 查询：人工智能有哪些应用？  (rewritten)
#      知识：[passages]
#      答：人工智能的应用包括...

# Clear memory for new conversation
agent.empty_memory()

Related Pages

• Heuristic:FlagOpen_FlagEmbedding_Warning_Deprecated_Search_Demo