Aron
Omnichannel AI Customer Support & Knowledge Engine
Built with a team for iliterate.ai.
Case Study
The problem
Support teams answer the same WhatsApp questions over and over, while the answers already sit in operational documents no customer ever opens. Script-based bots fail because real questions rarely follow the flow you prepared for.
Architecture
Docling extracts complex operational documents into text that preserves hierarchy; those chunks are embedded and stored as vectors in pgvector, inside the same PostgreSQL that holds operational data — no separate vector database to keep in sync. Redis holds conversation state, and a VLM module handles the image attachments customers send.
The hard part
Operational documents are full of tables and charts. Ordinary PDF parsers flatten them into sequential text, so a figure loses its row and column — and RAG then cites it confidently and wrongly. Docling was chosen specifically to preserve that structure when documents are split into chunks.
Outcome
The bot answers customer questions straight from company documents, with low-latency semantic search and answers grounded in sources rather than invented. Transfer receipts and complaint photos are recognised automatically without being escalated to a human.
System Architecture
Ingest
Docling
Extracts complex operational documents into hierarchy-preserving text.
Storage
pgvector · PostgreSQL
Chunks are embedded and stored as vectors, inside the same database as the operational data.
Redis
Holds conversation state.
AI path
VLMAI
Handles image attachments sent by customers.
Why it is shaped this way
pgvector lives in the same PostgreSQL as the operational data — there is no separate vector database to keep in sync.
Engineering Decisions
High-Fidelity Document Parsing (Docling)
Uses Docling to extract complex operational documents — tables, charts, layered PDFs — without losing the hierarchical structure of the information.
Retrieval-Augmented Generation (RAG)
pgvector inside PostgreSQL powers low-latency semantic search, letting the bot answer WhatsApp customer questions accurately, in context, and free of hallucination.
Multimodal Handling
A VLM module automatically recognises complaint photos, bank transfer receipts and documents customers send over WhatsApp.