kbfetch

Containerized RAG (Retrieval-Augmented Generation) Knowledge Base.

Technologies Used

• FastAPI: A modern, fast (high-performance) web framework for building APIs with Python 3.7+ based on standard Python type hints.
• Sentence Transformers: A Python framework for state-of-the-art sentence, text, and image embeddings. Specifically, all-MiniLM-L6-v2 is used for generating embeddings from document chunks.
• Qdrant: A vector similarity search engine and vector database. It stores vector embeddings and provides an API for similarity search.
• Ollama: A tool for running large language models locally. Phi-3 is used as the LLM for generating responses.
• Docker & Docker Compose: Used for containerizing the application components and orchestrating their deployment.

Getting Started

To get started with kbfetch, you need to have Docker and Docker Compose installed. The provided Makefile simplifies the process of building, running, and managing the application.

Available make Commands

• make auto: Automatically detects if a GPU is available and starts the appropriate services. This is the recommended way to start the application.
• make gpu: Starts the services with GPU acceleration.
• make cpu: Starts the services with CPU only.
• make build: Builds the Docker images.
• make down: Stops all running services.
• make logs: Shows the logs for all services.
• make clean: Stops and removes all containers, networks, and volumes associated with the application.

Quick Start

1. Clone the repository:

   git clone https://github.com/S4M8/kbfetch.git
   cd kbfetch

2. Build and start the services:

   make auto

   This command will detect if you have a GPU and start the appropriate services. If you want to force CPU or GPU mode, you can use make cpu or make gpu respectively.

Usage

Using the kbfetch CLI

A command-line interface (CLI) tool kbfetch is provided to simplify interaction with the kbfetch application.

Uploading Documents

To upload a document, use the upload command followed by the path to your document file:

kbfetch upload <path_to_your_document>

Querying the Knowledge Base

To query the knowledge base, use the query command followed by your query text:

kbfetch query "Your question here"

Planned Features

📄 Document Format Support

• JSON parsing and embedding
• YAML document support
• PDF content extraction (including OCR capabilities)
• Image processing and embedding (e.g., screenshots, scanned pages)
• CSV and Excel document ingestion
• XML document parsing
• TOML config file support

🔍 Search & Query Enhancements

• Semantic and intent-aware search
• Faceted filtering (type, tags, date, source, etc.)
• Full-text OCR for scanned documents and image-based PDFs

🧠 Metadata & Classification

• Automatic metadata extraction (e.g., author, date, keywords)
• Custom tagging system for documents
• NLP-driven auto-classification into topics or categories

🔗 UI, Workflow & Automation

• Web-based dashboard for browsing and managing knowledge base entries
• Rich document previews (syntax highlighting, PDF/image thumbnails, tree views)
• Commenting and collaborative annotations
• Bulk and recursive directory uploads
• Export content or search results to PDF, DOCX, or HTML
• Workflow automation (e.g., auto-tagging, alerts, custom triggers)

🔐 Security & Access Control

• Role-based access control
• Document-level permissions
• Audit logging and version history tracking for changes

🔧 Integrations & Extensibility

• API endpoints for external integration (e.g., Slack, GitHub, Zapier)
• Webhooks and event triggers (e.g., on document addition/modification)
• Pluggable embedding model support (e.g., OpenAI, Cohere, local models)

📊 Analytics & Reporting

• Usage and search analytics dashboard
• Document type breakdown reports
• Query success/failure rate tracking

🌐 Internationalization & Localization

• Multilingual document support
• Language detection and tagging

Profiling (Initial Version)

The following profiling was conducted on a machine with an Intel i9 CPU, 64GB of RAM, and an NVIDIA 3090 GPU. The llm_service was running in CPU mode.

Resource Usage

Service	State	CPU % (Peak)	Memory Usage (Peak)
rag_api	Idle	0.41%	2.077 GiB
	Upload	0.43%	2.113 GiB
	Query	0.40%	2.113 GiB
vector_db	Idle	0.24%	122.3 MiB
	Upload	0.38%	124.8 MiB
	Query	0.25%	123.2 MiB
llm_service	Idle	0.00%	12.94 MiB
	Upload	0.00%	12.94 MiB
	Query	~100%	3.998 GiB

Estimated Minimum Hardware Requirements

• RAM: 8 GB
• CPU: 4-core
• Storage: 10 GB

Note: The primary bottleneck is the llm_service during query operations. Using a supported GPU for the llm_service would significantly improve performance and reduce CPU load.