hArI — Document Intelligence System

Talk to your documents. Understand your data.

hArI is an AI-powered document intelligence system that lets you upload PDFs, CSVs, and Excel files — then have a natural conversation with their contents. It uses a RAG (Retrieval-Augmented Generation) pipeline for PDF semantic search and LLM-driven pandas code execution for structured data analysis.

---

Features

• PDF Chat — Semantic search over PDF content using ChromaDB + sentence-transformers
• CSV / Excel Analysis — Natural language queries converted to pandas operations via LLM
• Mixed File Mode — Upload both types together; hArI automatically routes each query to the right engine
• Streaming Responses — Token-by-token streaming with live cursor effect (ChatGPT-style)
• Conversation Memory — Maintains context across turns; summarizes older messages when the context window fills up
• Persistent Vector Store — PDF embeddings stored across sessions; no re-processing on reload
• Modular Architecture — Clean separation of UI, state, handlers, and core logic

---

Tech Stack

Layer	Technology
Language	Python 3.10+
UI	Streamlit
RAG Pipeline	ChromaDB (cosine similarity)
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
LLM	Groq API (llama-4-scout-17b, compound-beta-mini)
Data Analysis	Pandas + NumPy
PDF Parsing	PyMuPDF

---

Project Structure

hArI/
├── app.py                  # Entry point — page config + render orchestration
├── config.py               # All config constants (models, chunking, RAG, memory, limits)
│
├── core/
│   ├── __init__.py         
│   ├── file_processor.py   # File loading, type detection, PDF/CSV/Excel parsing
│   ├── memory.py           # Conversation context + summarization logic
│   ├── query_intent.py     # Query classifier — routes to PDF (RAG) or CSV engine
│   ├── responser.py        # Prompt builder + streaming Groq response handler
│   └── utils.py            # Shared utilities (strip_thinking, Groq client helpers)
│
├── rag/
│   ├── __init__.py         
│   ├── embedder.py         # Splits docs into chunks, generates embeddings
│   ├── retriever.py        # Embeds user query, retrieves top-k chunks from ChromaDB
│   └── vector_store.py     # ChromaDB client, collection management, deduplication
│
├── ui/
│   ├── __init__.py         
│   ├── styles.py           # Global CSS injection (dark theme, purple accents)
│   ├── state.py            # Session state init + clear_chat()
│   ├── handlers.py         # Ingest, query, reset, remove file pipelines
│   └── components.py       # All Streamlit render functions
│
├── prompts/
│   └── system_prompt.md    # AI identity + mode-specific rules (PDF, CSV, General)
│
├── data/
│   └── chroma_store/       # Persistent ChromaDB vector storage
│
├── .env                    # API keys (not committed)
├── pyproject.toml
└── uv.lock

---

Getting Started

Prerequisites

• Python 3.10+
• uv (recommended) or pip
• A Groq API key

Installation

# Clone the repo
git clone https://github.com/yourusername/hArI.git
cd hArI

# Install dependencies using uv
uv sync

# Or using pip
pip install -r requirements.txt

Environment Setup

Create a .env file in the project root:

GROQ_API_KEY=your_groq_api_key_here

Run

streamlit run app.py

---

How It Works

PDF Mode (RAG Pipeline)

Upload PDF
    │
    ▼
PyMuPDF extracts text
    │
    ▼
Text split into chunks (embedder.py)
    │
    ▼
Embeddings generated (all-MiniLM-L6-v2)
    │
    ▼
Stored in ChromaDB (vector_store.py)
    │
    ▼
User query → embed → retrieve top-k chunks (retriever.py)
    │
    ▼
Chunks + memory context injected into prompt (responser.py)
    │
    ▼
Groq streams answer token-by-token → rendered live in UI

CSV / Excel Mode (Direct LLM)

Upload CSV / Excel
    │
    ▼
Pandas DataFrame created (file_processor.py)
    │
    ▼
Schema + sample rows extracted as metadata
    │
    ▼
User query + schema → Groq generates pandas code
    │
    ▼
Code executed safely → result passed back to LLM
    │
    ▼
Groq streams formatted answer → rendered live in UI

Mixed Mode

When both file types are present, query_intent.py uses the LLM to classify whether each query is best answered by the PDF RAG engine or the CSV analysis engine — then routes accordingly.

---

Configuration

All tunable parameters live in config.py:

Parameter	Description
CHUNK_SIZE	Token size per text chunk for PDF splitting
CHUNK_OVERLAP	Overlap between consecutive chunks
TOP_K_RESULTS	Number of chunks retrieved per query
MEMORY_BUFFER_SIZE	Max messages before summarization triggers
MAX_FILE_SIZE_MB	Upload size limit per file
EMBEDDING_MODEL	Sentence-transformer model name
LLM_MODEL	Groq model identifier
ANALYSIS_MODEL	Groq model used for pandas code generation
SCORE_THRESHOLD	Minimum cosine similarity score for chunk retrieval

---

Known Limitations

• PDF support only (no .docx, .txt currently)
• CSV/Excel analysis depends on LLM-generated pandas code — complex queries may occasionally fail

---

Roadmap

• Streaming LLM responses in UI
• Modular ui/ folder architecture
• ChromaDB cosine similarity threshold filtering
• Add .docx and .txt support
• Multi-collection support (separate namespaces per session)
• Export chat history
• Confidence score display on retrieved chunks

---

License

MIT License. See LICENSE for details.

---

Live Demo → hArI

Built by Harsh Bhanushali