RAG Knowledge Assistant

A retrieval-augmented generation system built from the ground up to understand and experiment with the core components of RAG.

Overview

This project focuses on understanding and implementing RAG fundamentals before introducing higher-level frameworks or managed vector databases.

Current pipeline:

Documents
   ↓
Text Extraction
   ↓
Structure-Aware Chunking
   ↓
Embeddings
   ↓
Semantic Retrieval
   ↓
Retrieval Evaluation

The initial document used for development is a resume, but the ingestion and retrieval pipeline is being designed to support other document sources.

Current Progress

• PDF text extraction
• Text cleaning
• Structure-aware chunking
• Chunk metadata
• Persistent chunk storage
• Sentence Transformer embeddings
• Persistent embeddings
• Cosine-similarity retrieval
• Top-k semantic retrieval
• Initial retrieval evaluation
• Retrieval improvements
• Metadata filtering
• Hybrid search
• Reranking
• Vector database
• Context construction
• LLM generation
• Answer citations
• RAG evaluation
• FastAPI service
• Docker deployment
• CI/CD

Retrieval Baseline

The current retriever uses:

• all-MiniLM-L6-v2
• 384-dimensional embeddings
• normalized embeddings
• cosine similarity
• top-k retrieval

Initial evaluation:

Metric	Score
Recall@3	0.750
Precision@3	0.333
MRR	0.583

These results represent the initial baseline and will be used to measure improvements to the retrieval pipeline.

Architecture

                 ┌──────────────────┐
                 │     Documents    │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │    Ingestion     │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │     Chunking     │
                 │   + Metadata     │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │    Embeddings    │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │    Retrieval     │
                 │  + Similarity    │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │      Context     │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │       LLM        │
                 └────────┬─────────┘
                          ↓
                 ┌──────────────────┐
                 │ Answer + Sources │
                 └──────────────────┘

Project Structure

rag-knowledge-assistant/
├── app/
├── data/
│   ├── processed/
│   └── raw/
├── notebooks/
├── src/
│   ├── ingestion.py
│   ├── chunking.py
│   ├── structure_chunking.py
│   ├── semantic_chunking.py
│   ├── chunking_pipeline.py
│   ├── embedding.py
│   ├── retrieval.py
│   └── evaluate_retrieval.py
├── tests/
├── .env.example
├── .gitignore
├── README.md
└── requirements.txt

Tech Stack

• Python
• Sentence Transformers
• NumPy
• PyPDF
• Git / GitHub
• Linux

Planned additions include Qdrant, reranking models, FastAPI, Docker, pytest, and GitHub Actions.

Roadmap

Phase 1 — RAG Fundamentals

Build the core pipeline without relying on RAG frameworks.

Phase 2 — Retrieval

Experiment with:

• chunk size and overlap
• semantic search
• metadata filtering
• similarity thresholds
• hybrid retrieval
• reranking

Phase 3 — Evaluation

Measure retrieval quality using:

• Recall@K
• Precision@K
• MRR

Then evaluate generated answers for relevance, faithfulness, and citation correctness.

Phase 4 — Production

Add:

• FastAPI
• structured logging
• configuration management
• testing
• Docker
• GitHub Actions
• latency measurement
• retrieval and generation tracing

Phase 5 — Advanced RAG

Explore:

• query rewriting
• multi-query retrieval
• HyDE
• parent-child retrieval
• contextual compression
• conversational memory

Advanced techniques will be added only after the baseline system is understood and evaluated.