PKA: Personal Knowledge Assistant (Backend)

An AI-powered Retrieval-Augmented Generation (RAG) backend service that allows users to upload, index, and query their local files and codebases. This service acts as the brain for the PKA_Frontend application, providing contextual answers and precise file/code references from your personal workspace.

🚀 Project Overview

PKA is designed to turn your local project directory into a searchable knowledge base.

• Users upload files and entire folders into a designated workspace directory on their machine.
• The system indexes the content using a custom, code-focused chunking approach.
• Users can then query the workspace (e.g., "Where is the handleSendMessage function defined?").
• The assistant (Misty the Cat 🐈) answers with relevant information, code snippets, and provides the relative file path to the source.

🔗 Ecosystem Links

This repository is the backend service. To interact with it, you will need the frontend application:

• PKA Frontend: https://github.com/riddhika05/PKA_Frontend

🧠 Technical Deep Dive: Code-Focused RAG

This project focuses on a lightweight and intuitive approach to RAG, deliberately avoiding heavier dependencies like LangChain and Tree-sitter.

Custom Indexing & Chunking Logic

The core logic, housed in build_VectorStore.py, uses a custom chunking pipeline to maximize the quality of code retrieval:

1. Text Files (.md, .txt, etc.): A simple, intuitive chunking function (chunk_text_fallback) splits documents into fixed-size segments (e.g., 1500 characters) with overlap (e.g., 200 characters).
2. Code Files (.py, .js, .jsx, .c, etc.): A more sophisticated function (chunk_code_by_functions) utilizes language-specific regular expressions (regex) to identify the start of key logical units (functions, classes, components). This ensures that complete, coherent blocks of code are kept intact as single chunks, making the RAG system far more effective for answering code-related questions.

The indexed chunks are stored in a ChromaDB vector store, and embeddings are generated using a Sentence Transformer model.

🛠️ Key API Endpoints

The service is built with FastAPI and uses a local Ollama instance running Phi3:mini for the LLM component.

Endpoint	Method	Description
/setup_workspace	POST	Creates the local workspace directory on the user's machine.
/upload_and_index	POST	Handles file/folder uploads, saves them to the workspace, and immediately indexes the new content.
/index_workspace	POST	Triggers a re-index of all files currently in the workspace folder.
/delete_file	DELETE	Removes a file from both the local workspace directory and the ChromaDB vector store.
/reset_collection	POST	WARNING: Deletes all indexed data from the ChromaDB collection.
/collection_stats	GET	Returns statistics on the knowledge base (e.g., total chunks, unique files).
/ask	POST	The main RAG endpoint. Accepts a question, searches the vector store for context, and streams the LLM response.

⚙️ Getting Started (Backend)

Prerequisites

• Python 3.8+
• A running local LLM instance (e.g., Ollama running phi3:mini).

Installation

1. Clone this repository.
2. Install dependencies:

   pip install -r requirements.txt

Running the Server

Start the API server using Uvicorn on port 8000:

uvicorn main:app --reload --host 0.0.0.0 --port 8000