Pramana

Pramana is an explainable multimodal RAG system for high-stakes document QA. It is designed for reports, board documents, financial decks, and visually dense PDFs where important facts may appear in normal text, tables, diagrams, charts, infographics, or page layout.

The production pipeline lives in service/src/powermind_rag and is API-first by default: NVIDIA-hosted embeddings, Mistral OCR, Gemini/NVIDIA VLM page understanding, Gemini CRAG relevance grading, and NVIDIA/Gemini final visual fallback.

Pramana also runs as an MCP server. See the MCP integration guide and environment format.

What Pramana Does

• Ingests PDF documents into local JSON records under service/storage.
• Extracts normal PDF text.
• Runs Mistral OCR for table and layout-heavy page content.
• Runs visual page understanding during ingestion:
  • Ingestion VLM order: Gemini first, NVIDIA Phi fallback.
  • Failed pages are recorded in visual_page_failures.json.
• Embeds chunks with NVIDIA nvidia/nv-embed-v1.
• Retrieves with BM25, keyword search, dense FAISS search, and Reciprocal Rank Fusion.
• Applies CRAG relevance grading before answer generation.
• Generates grounded answers with citations like [p3:c12].
• If the normal pipeline returns Not found in the document., performs one final page-image VLM check:
  • Query fallback order: NVIDIA Phi first, Gemini fallback.

Project Links

• Presentation
• Architecture Diagram
• Frontend Docs
• Backend Docs
• AI Workflow Docs
• MCP Integration
• Environment Format
• Docker Guide - Start here for containerization and Docker Hub

Docker (Containerization)

Pramana RAG is fully containerized for reproducible deployment across systems.

Quick start:

# Build image (~1.4GB, excludes large model files and generated data)
docker build -f Dockerfile.service -t pramana-rag:1.0.0 .

# Run with compose
cp .env.example .env.docker  # Add API keys
docker compose -f docker-compose.service.yml run --rm service \
  python -m powermind_rag.cli ask "What is this document?"

# Push to Docker Hub
docker tag pramana-rag:1.0.0 yourusername/pramana-rag:1.0.0
docker push yourusername/pramana-rag:1.0.0

Features:

• Multi-stage builds with a small runtime image.
• Volumes for models, input PDFs, and generated storage.
• Secrets passed at runtime, never baked into the image.
• Docker Compose orchestration with CLI and MCP modes.
• Docker Hub publishing workflow.

See DOCKER_README.md for full documentation: building, running CLI/MCP, volume management, publishing to Docker Hub, and troubleshooting.

Current Architecture

PDF
 |-- normal text extraction
 |-- page rendering
 |-- Mistral OCR
 |-- Gemini page analysis -> NVIDIA Phi fallback
 |
chunk factual propositions
 |
NVIDIA nv-embed-v1 embeddings
 |
service/storage/<document_id>/text_records.json
 |
query
 |-- BM25
 |-- keyword
 |-- NVIDIA dense query embedding + FAISS
 |-- RRF
 |-- Gemini CRAG relevance grading
 |-- grounded answer generation
 |-- final VLM page fallback when not found

Repository Layout

PowerMind/
|-- powermind_rag/                  # repo-root shim so python -m uses service source
|-- service/
|   |-- src/powermind_rag/          # production Pramana RAG package
|   |-- data/                       # input PDFs
|   |-- storage/                    # generated text records and page analysis
|   |-- tests/
|   |-- README.md
|-- backend/                        # compatibility FastAPI app
|-- frontend/                       # Next.js UI
|-- scripts/                        # runners and smoke scripts
|-- Presentation.pdf
|-- Architecture.jpeg

The old backend RAG v2/LangGraph/Pinecone path has been removed. The main runnable pipeline is the service package.

Environment

The repo-root .env is loaded automatically. Important values:

POWERMIND_STORAGE_DIR="D:\PowerMind\service\storage"
POWERMIND_DEVICE="api"

NVIDIA_KEY="..."
NVIDIA_EMBEDDING_MODEL="nvidia/nv-embed-v1"
NVIDIA_VLM_MODEL="microsoft/phi-4-multimodal-instruct"
NVIDIA_GENERATION_MODEL="microsoft/phi-4-multimodal-instruct"
NVIDIA_VLM_BASE_URL="https://integrate.api.nvidia.com/v1"

MISTRAL_API_KEY="..."
MISTRAL_SERVER_URL="https://api.mistral.ai"

GEMINI_API_KEY_1="..."
GEMINI_API_KEY_2="..."
GEMINI_API_KEY_3="..."
GEMINI_API_KEY_4="..."
GEMINI_API_KEY_5="..."
GEMINI_API_KEY_6="..."
GEMINI_RELEVANCE_MODEL="gemini-2.5-flash"

POWERMIND_ENABLE_VISUAL_UNDERSTANDING="true"
POWERMIND_VISUAL_UNDERSTANDING_PROVIDER="gemini_nvidia"
POWERMIND_IMAGE_PROVIDER="nvidia_gemini"
POWERMIND_ENABLE_QUERY_VLM_FALLBACK="1"
POWERMIND_RELEVANCE_PROVIDER="gemini"
POWERMIND_ALLOW_LEXICAL_CRAG_FALLBACK="0"

POWERMIND_ENABLE_COLPALI_VISUAL_INDEX="0"

POWERMIND_ENABLE_COLPALI_VISUAL_INDEX is optional and off by default. Turning it on loads the optional ColPali/byaldi visual index. Keep it off for the API-first deployment.

Install

From the service folder:

cd D:\PowerMind\service
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r requirements.txt
python -m pip install -e . --no-deps

From the repo root, python -m powermind_rag.cli ... uses service/src/powermind_rag through the root shim.

Verify module resolution:

cd D:\PowerMind
python -c "import importlib.util; print(importlib.util.find_spec('powermind_rag.cli').origin)"

Expected:

D:\PowerMind\service\src\powermind_rag\cli.py

Ingest

Single PDF:

cd D:\PowerMind
python -m powermind_rag.cli ingest "D:\PowerMind\service\data\AEL_Earnings_Presentation_Q2-FY26_copy.pdf"

All PDFs in service/data:

cd D:\PowerMind
python -m powermind_rag.cli ingest-dir "D:\PowerMind\service\data"

Ingestion writes:

service/storage/<document_id>/text_records.json
service/storage/<document_id>/visual_page_analysis.json
service/storage/<document_id>/visual_page_failures.json

Ask

cd D:\PowerMind
python -m powermind_rag.cli ask "What is the Adani Family's equity stake in AEL as shown in the portfolio structure diagram?" --show-timings

Batch Questions

cd D:\PowerMind
python -m powermind_rag.cli ask-batch --output service\outputs\qa_results.md --json-output service\outputs\qa_results.json

Notes

• No local model is required in the default deployment path.
• Optional local components remain in the codebase for earlier experiments and future toggles.
• Generated Python caches are intentionally not kept.
• Stored embeddings are auto-migrated when the configured embedding model changes.