The foundational data lakehouse + internal logic engine of JOY Automart. ONI does not make business decisions. Its only job: ingest unstructured chaos (clicks, supplier PDFs, fitment logs) and weave it into a flawless, queryable narrative. JOYADH asks. ONI answers. The army acts.
Maintain the four JOY Automart narratives — Vehicle & Fitment · Customer Behaviour · Supply Truth · Margin Reality. Provide every other agent a single, instantly-queryable Truth™ on demand.
Clean dirty supplier APIs · categorise customer behaviour · normalise scrapes into pricing graphs.
When JOYADH needs a decision, ONI returns a unified data packet linking past purchases, current warehouse stock and weather context.
Hundreds of thousands of parts mapped to exact Make · Model · Year · Engine Trim.
Tracks every hover, abandoned cart, search query. Builds the unified profile that ZIFA personalises against.
Logs every JASMINE movement. Calculates the real cost of a brake pad factoring in hidden shipping delays.
Bengali ↔ English semantic match so a slang query 'চিমটা' returns 'Lower Control Arm'.
ONI is the reason the system learns. Every ZIFA ad result, every ARES combat brief, every JASMINE fulfilment outcome — ONI records it. The next JOYADH decision is smarter because of it. The architecture compounds.
ONI is strictly observational. Even if the data clearly shows a mispriced transmission fluid, ONI cannot change the price. It must wait for JOYADH to read the data and order ARES or ZIFA to act.
| Layer | Tech | Purpose |
|---|---|---|
| Relational anchor | PostgreSQL + pgvector / MariaDB Vector | ACID-compliant SKU counts · VIN fitment · supplier POs · ledgers |
| Vector brain | pgvector · Pinecone / Qdrant | Semantic similarity for behavioural profiles · RAG memory |
| Lakehouse | Databricks · Snowflake · S3 + Iceberg | Cheap dump for raw HTML scrapes · supplier PDFs |
| Hybrid queries | Single SQL · vector + relational | Find semantically relevant parts WHERE inventory>0 AND margin>15% |