kbbridge 0.1.4

MCP server for intelligent knowledge base search and retrieval with Dify integration

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:

Backend	Status	Notes
Dify	Supported	Currently available
Others	Planned	Additional backends coming soon

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: Kubernetes with Helm (Recommended for Production)

For production Kubernetes deployments, use the Helm chart:

# Build and push Docker image to your registry
docker build -t your-registry/kbbridge:0.1.0 .
docker push your-registry/kbbridge:0.1.0

# Install with Helm
helm install kbbridge ./helm/kbbridge \
  --set image.repository=your-registry/kbbridge \
  --set image.tag=0.1.0 \
  --set env.RETRIEVAL_API_KEY=your-key \
  --set env.LLM_API_TOKEN=your-token

See helm/kbbridge/README.md for detailed configuration options.

Option 2: 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

When to use what:

• Helm/Kubernetes: Production clusters, multi-node deployments, scaling, orchestration, high availability
• Docker/Docker Compose: Local development, quick testing, simple single-node deployments, CI/CD pipelines

Note: If you're deploying to Kubernetes, you only need Helm. Docker Compose is optional and primarily for local development convenience.

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

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 mcp import ClientSession

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

asyncio.run(main())

With Custom Instructions

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

With Quality Reflection

await session.call_tool("assistant", {
    "dataset_id": "dataset_id",
    "query": "What are the safety protocols?",
    "reflection_mode": "standard",  # "off", "standard", or "comprehensive"
    "reflection_threshold": 0.75,
    "max_reflection_iterations": 2
})

Reflection Modes

• off: No reflection (fastest)
• standard (default): Answer quality evaluation only
• comprehensive: Search coverage + answer quality evaluation

Reflection evaluates answers on:

• Completeness (30%): Does the answer fully address the query?
• Accuracy (30%): Are sources relevant and correctly cited?
• Relevance (20%): Does the answer stay on topic?
• Clarity (10%): Is the answer clear and well-structured?
• Confidence (10%): Quality of supporting sources?

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