DOTNET-RAG-Document-QA

Author: Krish Nangia

A .NET 8 console application implementing a RAG pipeline for document-based question answering.
This code ingests plain text documents, generate embeddings for semantic search, and query an LLM to get answers grounded strictly in the provided documents.

---

Overview

This application is designed to demonstrate a simple RAG workflow:

1. Document Ingestion – Read and chunk text documents.
2. Vector Embedding – Generate embeddings for each chunk using Gemini.
3. Vector Store – Store chunk embeddings with metadata for retrieval.
4. Query Handling – Accept user questions via the console.
5. Similarity Search – Retrieve the most relevant chunks using cosine similarity.
6. LLM Answer Generation – Send retrieved context + user question to LLM to produce a grounded answer.

---

Step-by-Step Flow

1. Program Startup

• Loads FileVectorStore from the configured VECTOR_STORE_LOCATION.
• If the folder does not exist, initializes a new store.

2. Document Ingestion

• Scans DOCUMENTS_FOLDER_PATH for .txt files.
• Checks whether files were previously ingested by comparing existing vector store chunks.
• Already ingested files are skipped to avoid duplication.
• Currently, ingesting all documents in the Stories directory from scratch takes around 6 minutes.

3. Chunking Documents

• Reads each file’s content.
• Uses SemanticSlicer to split the document into smaller chunks (e.g., 50–100 words per chunk).

4. Embedding Generation

• For each chunk, calls GeminiEmbeddingGenerator to produce a vector embedding.
• Stores metadata for each chunk:
  • SourceFile – the original filename
  • ChunkIndex – position of the chunk within the file

5. Vector Store Update

• Adds each chunk and its embedding to FileVectorStore.
• Implements a delay (delayMs) between API calls to avoid rate limiting.

6. User Query Handling

• Waits for console input from the user.
• Accepts natural language questions.
• Type exit to quit.

7. Similarity Search

• Computes similarity (cosine similarity) between query embedding and stored chunk embeddings.
• Retrieves the top N most relevant chunks (default: 5).

8. Context Construction

• Combines the retrieved chunks into a single context string.
• Optionally includes metadata such as SourceFile and ChunkIndex for debugging.

9. LLM Query & Answer Generation

• Sends the context and question as a prompt to the LLM (LLMService).
• Receives an answer grounded strictly in the provided context.

10. Output Answer

• Prints the answer to the console.
• Loops back to accept more queries until the user exits.

---

Features

• Incremental document ingestion.
• Chunk-based vector embeddings for precise semantic retrieval.
• Top-N similarity search for relevant context.
• LLM-based answer generation grounded in retrieved chunks.
• Simple console interface with safe exit handling.

---

Approach & Assumptions

• Unique filenames are used for ingestion detection. If a file is modified but the name is unchanged, the system will not re-ingest it.
• The vector store preserves all previous embeddings; old chunks remain even if files are updated.
• Works only with plain text documents. PDFs or Word documents require preprocessing into .txt.
• Delays (delayMs) are used to avoid hitting rate limits.
• Metadata such as SourceFile and ChunkIndex is optional but useful for debugging and context reference.

---

Setup & Installation

git clone https://github.com/knangia04/DOTNET-RAG-Document-QA.git
cd DOTNET-RAG-Document-QA
dotnet restore
dotnet build

• Place your documents in the Documents folder and set GlobalSettings.DOCUMENTS_FOLDER_PATH in the GlobalSettings.cs.
• Set GlobalSettings.VECTOR_STORE_LOCATION for storing chunk embeddings.

---

Usage

Run the application:

dotnet run

• The program will ingest new documents and generate embeddings.
• After ingestion, you can enter questions interactively.
• Type exit to quit the program safely.