A step-by-step tour of the architecture powering grounded, workspace-isolated Oracle ERP answers.
Upload Oracle PDFs, Word docs, or just ask a question. The RAG ingestion pipeline extracts text, redacts sensitive data, chunks content, and generates NVIDIA NIM embeddings for vector search.
Each question runs against a workspace-isolated vector store. Relevant chunks surface from both your uploaded documents and Oracle public knowledge, ranked by similarity.
A specialist Google ADK agent receives the grounded context and your question, then drafts an answer, SQL query, PL/SQL script, or solution blueprint — with source citations.
Read queries execute directly against your Oracle instance. Write operations are allowlist-gated by default — the agent drafts them and you approve before execution. Production writes require explicit opt-in.
Answers, blueprints, and SQL appear in the IDE. Download formatted exports as PDF or Excel for client sign-offs, audit trails, or your knowledge library.
The self-learning engine captures corrections and builds project memory over time. Next time you ask a similar question, the agent already knows your ERP setup and preferences.
Everything listed here is verified against the codebase — no marketing fluff.
Frontend
Next.js 16 App Router, React 19, Tailwind CSS, ShadCN UI
AI Orchestration
Google ADK 2.0 Graph, 26 Specialist Agents, Supervisor Swarm
LLM Models
Gemini 2.0 Flash (primary), configurable via AI Provider settings
Embeddings
NVIDIA NIM nv-embedcode-7b-v1 (verified in NVIDIA catalog)
Vector Store
Firestore-based workspace-scoped RAG with cosine similarity
Database Access
Oracle EBS R12 (ISG REST, PL/SQL wrapper, oracledb) & Fusion REST
Payments
Razorpay (INR orders, webhook verification, invoice generation)
Auth & Security
Firebase Auth (Google SSO only), AES-256-GCM encryption, session revocation
Hosting
Firebase App Hosting + Cloud Functions v2
The live demo on the homepage runs against the real API. Sign in with Google to get your own workspace.
