Vectorless RAG, Semantic Document Search | Try Live Demo

Real-time pipeline for retrieving and analyzing documents using a Custom BM25 Page-Index retrieval engine in Python — with automated text chunking, TF-IDF ranking, and a high-speed LLaMA generation fallback (Groq).

Results at a Glance

	Vectorless (BM25)	Vector DB (Embeddings)
Approach	TF-IDF / Local BM25	Semantic Embeddings
Indexing Speed	Near Instant	Slower (API Dependency)
Storage Size	Minimal (In-Memory)	High (Vector Storage)
Robustness	High (Exact Keywords)	High (Contextual)
Generation	Groq LLaMA Inference	External API Inference

BM25 Page-Index Engine

• Search Config: k1=1.5, b=0.75
• Retrieval Engine: Custom Inverted Index
• Preprocessing: Tokenization, Stopwords Removal, Term Freq.
• Chunking Strategy: By Page (.pdf) or 2000c Blocks (.txt)

Groq LLaMA Model (Generator)

• Model Baseline: llama-3.3-70b-versatile
• Confidence Constraint: Strict Grounding (answers only using context)
• Latency Setup: Very Fast

---

How It Works

• Observation (Document Input): PDF or TXT files uploaded via Streamlit, text extracted and chunked page-by-page.
• Action/Prediction (Retrieval): Custom BM25 search engine tokenizes the user query and calculates TF-IDF scores to retrieve the top-4 relevant pages.
• Key preprocessing components: Automated page extraction—rag_engine.py builds an inverted index over the extracted pages, dropping common stopwords and optimizing search dynamically.
• Termination/Fallback: The LLM receives the constructed context and generates a plain-English answer grounded purely in the document, explicitly citing sources. If no context correlates, it refuses the answer to prevent hallucinations.

---

Setup

pip install streamlit requests pypdf

Vectorless_RAG/
├── app.py               # Streamlit application logic & chat interface
├── rag_engine.py        # Vectorless retrieval, tokenization, & inverted index
└── .streamlit/
    └── config.toml      # Configuration settings for custom UI themes

---

Usage

# 1. Set your Groq API Key
$env:GROQ_API_KEY="gsk_..."  # Windows PowerShell
export GROQ_API_KEY="gsk_..." # Linux/macOS

# 2. Launch the Streamlit App
streamlit run app.py

---

Key Design Decisions

• Custom BM25 Engine — completely vectorless architecture using robust token transformations and inverted index, achieving high-speed term-based extraction without massive dependencies.
• Strict Grounding criteria — ensures hallucinations are strictly classified as "not present" for better contextual accuracy.
• Automated page-level chunking — rag_engine.py parses PDFs methodically maintaining exact page structures, so source citations are flawlessly accurate instead of drifting boundaries.
• Groq LLaMA Deployment — provides an immediate, low-latency LPU deployment alternative without requiring heavy cost or local tensors.

---

License

MIT — see Streamlit and Groq for dependencies.