egpivo/kb-bridge

KB-Bridge

A Model Context Protocol (MCP) server for intelligent knowledge base search and retrieval with support for multiple backend providers.

Installation

pip install kbbridge

Quick Start

Configuration

Create a .env file with your retrieval backend credentials:

# Required - Retrieval Backend Configuration
RETRIEVAL_ENDPOINT=https://api.dify.ai/v1  # Example: Dify endpoint
RETRIEVAL_API_KEY=your-retrieval-api-key
LLM_API_URL=https://your-llm-service.com/v1
LLM_MODEL=gpt-4o
LLM_API_TOKEN=your-token-here

# Optional
RERANK_URL=https://your-rerank-api.com
RERANK_MODEL=your-rerank-model

Supported Backends:

See env.example for all available configuration options.

Running the Server

# Start server
python -m kbbridge.server --host 0.0.0.0 --port 5210

# Or using Makefile (if available)
make start

Server runs on http://0.0.0.0:5210 with MCP endpoint at http://0.0.0.0:5210/mcp.

Deployment Options

Option 1: Docker (Local Development / Simple Deployments)

For local development or simple single-container deployments:

# Build the image
docker build -t kbbridge:latest .

# Run with environment variables
docker run -d \
  --name kbbridge \
  -p 5210:5210 \
  --env-file .env \
  kbbridge:latest

For production deployments, use container orchestration platforms like Kubernetes with your preferred deployment method.

Features

• Backend Integration: Extensible architecture supporting multiple retrieval backends
• Multiple Search Methods: Hybrid, semantic, keyword, and full-text search
• Quality Reflection: Automatic answer quality evaluation and refinement
• Custom Instructions: Domain-specific query guidance

Workflow

KB-Bridge follows a multi-stage pipeline to ensure high-quality answers:

flowchart LR
    Start([User Query]) --> Preprocess[Query Preprocessing<br/>Rewriting & Understanding]

    Preprocess --> FileDiscovery[File Discovery<br/>Find Relevant Files]

    FileDiscovery --> Search[Search Stages]

    Search --> Direct[Direct Approach<br/>Simple Retrieval]
    Search --> Advanced[Advanced Approach<br/>File-level Processing]

    Direct --> Candidates
    Advanced --> Candidates[Collect Candidates]

    Candidates --> Synthesis[Answer Synthesis<br/>Rerank & Format]

    Synthesis --> Reflection{Reflection<br/>Enabled?}
    Reflection -->|Yes| Reflect[Quality Check<br/>& Refinement]
    Reflection -->|No| Final
    Reflect --> Final([Final Answer])

    style Start fill:#e1f5ff
    style Final fill:#c8e6c9
    style FileDiscovery fill:#fff9c4
    style Direct fill:#fff9c4
    style Advanced fill:#fff9c4
    style Reflect fill:#ffccbc
    style Synthesis fill:#e1bee7

Stage Details

Query Preprocessing (Optional)

• Query Rewriting: LLM-based expansion/relaxation to improve recall
• Query Understanding: Extract intent and decompose complex queries

File Discovery

• Semantic search to identify relevant files (recall-focused)
• Optional quality evaluation with automatic search expansion if quality is low

Search Stages (Parallel)

• Direct Approach: Simple query → retrieval → answer extraction (fallback)
• Advanced Approach: File-level processing with content boosting for precision

Answer Synthesis

• Rerank candidates by relevance (if reranking service available)
• Combine and deduplicate using LLM

Quality Reflection (Optional)

• Evaluate answer quality and refine if needed (up to max_iterations)

Implementation Status

The orchestrator (DatasetProcessor) currently implements stages 1-3, 5-8. File Discovery Quality Evaluation (stage 4) is implemented but not yet integrated into the pipeline. See .doc/FILE_DISCOVERY_EVALUATION_CONFIG.md for details.

Available Tools

• assistant: Intelligent search and answer extraction from knowledge bases
• file_discover: Discover relevant files using retriever + optional reranking
• file_lister: List files in knowledge base datasets
• keyword_generator: Generate search keywords using LLM
• retriever: Retrieve information using various search methods
• file_count: Get file count in knowledge base dataset

Usage Examples

Basic Query

import asyncio
from fastmcp import Client


async def main():
    async with Client("http://localhost:5210/mcp") as client:
        result = await client.call_tool(
            "assistant",
            {
                "resource_id": "resource-id",
                "query": "What are the safety protocols?",
            },
        )
        print(result.content[0].text)

asyncio.run(main())

With Custom Instructions

await client.call_tool("assistant", {
    "resource_id": "hr_dataset",
    "query": "What is the maternity leave policy?",
    "custom_instructions": "Focus on HR compliance and legal requirements."
})

With Query Rewriting

await client.call_tool("assistant", {
    "resource_id": "resource-id",
    "query": "What are the safety protocols?",
    "enable_query_rewriting": True  # Enables LLM-based query expansion/relaxation
})

With Document Filtering

await client.call_tool("assistant", {
    "resource_id": "resource-id",
    "query": "What are the safety protocols?",
    "document_name": "safety_manual.pdf"  # Limit search to specific document
})

Integration with Dify

You can plug KB-Bridge into a Dify Agent Workflow instead of calling MCP tools directly:

1. Configure MCP Connection
   • MCP server URL: http://localhost:5210/mcp
   • Add auth headers: X-RETRIEVAL-ENDPOINT, X-RETRIEVAL-API-KEY, X-LLM-API-URL, X-LLM-MODEL
2. Create an Agent Workflow
   • Add an “MCP Tool” node
   • Select tool: assistant
   • Map workflow variables to resource_id, query, and other tool parameters
3. Run Queries
   • User input → Agent → MCP assistant tool → Structured answer with citations

Development

# Install development dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/

# Format code
black kbbridge/ tests/

# Lint code
ruff check kbbridge/ tests/

License

Apache-2.0