OpsMind is an open source tool that connects to a SQL database, ingests PDF documents, and lets you ask questions in natural language. It uses a local LLM via Ollama — no cloud, no API keys, no data leaves your machine.
A Streamlit app with 7 tabs. Connects to SQLite (demo) or SQL Server (production).
Type a question. The LLM converts it to SQL, runs it against your database, and explains the result. Pre-built queries handle the 10 most common questions without LLM generation.
Upload PDFs. ChromaDB indexes them into vector embeddings. Ask questions like "What is the allergen procedure?" and get the relevant paragraph with source citation.
Production output, waste, yield, and orders visualised with Plotly charts. KPIs update from live database queries. Waste cost shown in GBP.
Batch traceability by code. Temperature excursion detection. Allergen matrix generation. Compliance score calculation.
Checks for: yield drops vs 30-day average, temperature excursions, overtime breaches (Working Time Regulations), expiring stock, order shortfalls.
Upload a spreadsheet, ask a question about it. Uses Pandas for analysis. The LLM summarises the results.
Architecture diagram.
# User types a question
"How much salmon did we process this week?"
# Step 1: Query Library checks 10 pre-built patterns
Match found → uses tested SQL (no LLM needed)
# Step 2: If no match → Schema Registry selects relevant tables
150 tables → only 4 sent to LLM prompt
# Step 3: LLM generates SQL (Phi3 Mini via Ollama)
SELECT pr.name, SUM(p.finished_output_kg) ...
# Step 4: SQLAlchemy executes query on your database
# Step 5: LLM explains the result in plain English
# Step 6: Plotly auto-generates a chart if data is numericAll dependencies are free and open source.
| Component | Tool | Purpose |
|---|---|---|
| LLM | Ollama (Phi3 Mini / Mistral 7B) | Natural language to SQL, result explanation |
| Database | SQLAlchemy | Supports SQLite (demo) and SQL Server (production) |
| Vector Search | ChromaDB + sentence-transformers | PDF document search (RAG) |
| UI | Streamlit | Dashboard, chat, charts |
| Charts | Plotly | Interactive production and waste visualisation |
| Data | Pandas | Excel/CSV analysis |
| Config | YAML | Schema registry for mapping table names |
How it handles large databases (100+ tables).
The LLM cannot hold 147 table definitions in its prompt. The schema registry maps business domains to specific tables. When you ask about "orders", only order-related tables are sent to the LLM — not all 147.
# schema.yaml — map your actual table names
traceability:
tables:
ProductionBatch: BatchID, BatchNo, ProductCode, ProductionDate
RawMaterialIntake: IntakeID, ProductCode, BatchNo, SupplierCode
Products: ProductCode, ProductName, Species, Allergens
SalesOrders: OrderID, CustomerCode, QuantityKg, DeliveryDate
# 7 domains: traceability, production, orders, temperature, staff, stock, compliance
# Each domain maps to 2-5 tables instead of 14736 pytest tests. All passing.
$ python -m pytest tests/test_core.py -v
tests/test_core.py::TestConfig::test_config_loads PASSED
tests/test_core.py::TestSQLDialect::test_days_ago_sqlite PASSED
tests/test_core.py::TestSchemaRegistry::test_detect_domain_traceability PASSED
tests/test_core.py::TestDatabase::test_query_returns_dataframe PASSED
tests/test_core.py::TestCompliance::test_trace_batch_returns_dict PASSED
tests/test_core.py::TestAlerts::test_check_all_alerts_returns_list PASSED
tests/test_core.py::TestWastePredictor::test_predict_waste PASSED
tests/test_core.py::TestSQLAgentSafety::test_blocks_insert PASSED
tests/test_core.py::TestDocSearch::test_search_returns_list PASSED
... 27 more tests
36 passed, 0 failedRequires Python 3.11+ and Ollama.
# 1. Install Ollama from https://ollama.com
ollama pull phi3:mini
# 2. Clone and install
git clone https://github.com/Pawansingh3889/OpsMind.git
cd OpsMind
pip install -r requirements.txt
# 3. Create demo database (662 production runs, 451 orders, 3600 temp logs)
python scripts/seed_demo_db.py
python scripts/ingest_documents.py
# 4. Run
streamlit run app.py
# 5. Connect to SQL Server (optional)
# Set environment variable:
OPSMIND_DB=mssql+pyodbc://user:pass@server/database?driver=ODBC+Driver+17+for+SQL+Server| LLM accuracy | Phi3 Mini generates incorrect SQL for complex queries. Pre-built query library handles the 10 most common questions reliably. |
| Response time | 10-25 seconds per query on 16GB RAM. The LLM is the bottleneck, not the database. |
| No authentication | Streamlit app has no login system. Intended for single-user or trusted network use. |
| Read-only | Only SELECT queries are allowed. Cannot write to or modify the database. |
| Demo data | Ships with synthetic data. Not tested on production databases yet. |
OpsMind/
├── app.py # Streamlit app (7 tabs)
├── config.py # Configuration
├── schema.yaml # Table mapping for large databases
├── modules/
│ ├── sql_agent.py # NL to SQL + pre-built query library
│ ├── query_library.py # 10 tested SQL patterns
│ ├── schema_registry.py # Domain-to-table mapping
│ ├── sql_dialect.py # SQLite/SQL Server abstraction
│ ├── database.py # SQLAlchemy connection layer
│ ├── doc_search.py # ChromaDB RAG
│ ├── compliance.py # Traceability, allergens, audit
│ ├── alerts.py # Anomaly detection (5 alert types)
│ ├── waste_predictor.py # Yield trends, waste analysis
│ └── llm.py # Ollama connection
├── tests/
│ └── test_core.py # 36 pytest tests
├── scripts/
│ ├── seed_demo_db.py # Demo database generator
│ └── ingest_documents.py # Sample document loader
└── requirements.txt