Innsight-AI is a modular Hybrid Business Intelligence AI system that combines structured analytics, business formula computation, and document intelligence into one unified architecture.
It runs fully offline using a local LLM (Ollama) and persistent vector storage, with no cloud APIs or external dependencies.
- Sum / Total
- Average / Mean
- Min / Max
- Count
- Group-by (Department, Vendor, Category, Date)
- Automatic graph rendering (Bar & Line)
- Dynamic column detection (no hardcoding)
Supports:
- RevPAR
- ADR
- Occupancy Rate
- ROI
- Profit Margin
Built using a registry-based extensible formula system.
- PDF semantic search
- Sentence-based chunking
- Embeddings using SentenceTransformers
- Persistent ChromaDB vector storage
- Context-grounded LLM responses
Automatically routes user queries to:
- Formula Engine
- Analytics Engine
- Vector Search + LLM
- Fallback LLM
Structured queries bypass the LLM entirely for performance.
Streamlit UI (app.py)
↓
Query Router (rag_chat.py)
↓
┌───────────────┬────────────────┬────────────────┐
Analytics Formula Engine Vector Engine
(analytics.py) (business_formulas.py) (vector_store.py)
↓
Document Registry (SQLite)
↓
Sync Engine (sync_engine.py)
Strict separation of responsibilities across layers.
project_root/
│
├── app.py
├── rag_chat.py
├── analytics_engine.py
├── business_formulas.py
├── vector_store.py
├── sync_engine.py
├── document_registry.py
│
└── storage/
├── documents/
└── chroma_db/
| Layer | Technology |
|---|---|
| UI | Streamlit |
| Analytics | Pandas |
| Graphing | Matplotlib |
| Embeddings | SentenceTransformers |
| Vector DB | ChromaDB (Persistent) |
| LLM | Ollama (tinyllama / phi) |
| Registry | SQLite |
| Backend | Pure Python |
No:
❌ LangChain
❌ FAISS
❌ OpenAI API
❌ Cloud services
Fully offline after model download.
Ensure Python 3.10+ and required packages are installed.
In a separate terminal:
ollama serveInstall model if not already installed:
ollama pull tinyllamastreamlit run app.py- sum amount by department
- average profit by vendor
- count transactions by category
- calculate revpar
- what is occupancy rate
- what is the refund policy?
- explain data mining
- Detects newly added files
- Prevents duplicate indexing using SHA256 hashing
- Removes vectors when files are deleted
- Maintains persistent SQLite registry
- Lazy loading of embedding model
- Lazy initialization of ChromaDB
- Controlled LLM invocation
- Batch embedding for efficiency
- No unnecessary re-indexing
- Structured queries avoid LLM calls
- Fully local inference
Optimized for low-resource environments.
After:
- Installing Ollama
- Downloading embedding model
The system works completely offline.
No internet required.
- Hotel Industry BI
- Financial Reporting
- Enterprise Analytics
- Operational Monitoring
- Hybrid AI Intelligence Platforms
- Hybrid Structured + Unstructured Intelligence
- Enterprise-style modular architecture
- Persistent vector memory
- Auto graph generation
- Local LLM integration
- Zero cloud dependency
- Smart routing for performance
- Multi-CSV dataset selector
- Semantic column matching
- Confidence scoring
- Role-based access control
- REST API layer
- Multi-model routing
- Dashboard-style UI upgrade
Aryan
Software Architecture & Hybrid AI Systems Enthusiast
MIT License