Four microservices powering retrieval, generation, configuration and memory as a knowledge agent inside a larger agentic system.
We do not publish client names
Teams across several departments were losing time searching an internal wiki and downloading files to answer questions their own documentation already covered. The platform had to serve multiple tenants with isolated indices and slot into a wider ecosystem of notification, email and reporting agents.
Four services: ingestion, generation, config and memory. Ingestion runs event-driven Celery tasks that extract and chunk documents, embed them and index into per-tenant OpenSearch indices. Generation plans queries against conversation history and a glossary, expands each into up to five keyword and semantic variants, retrieves with a choice of BM25-only, semantic-only or hybrid strategies, and reranks with reciprocal rank fusion or one of three cross-encoder models. A critique node scores and refines answers over configurable iterations. Redis provides semantic caching, and results stream to clients over Server-Sent Events.
A working multi-tenant knowledge agent. The memory service retains question, answer and node-level execution data to support multi-turn reasoning, and semantic caching removes redundant model calls.
Atlas turns scattered documents, meeting notes, and transcripts into a connected map of the people, projects, and facts inside them. Ask a question in plain language and Atlas…
Read the case study Atlas — Your company’s evergrowing knowledge, mappedConcurrent document processing across Go and Python services, with retrieval-augmented generation over legal documents.
Our own product, designed and built end to end: QR check-in with no hardware, GST-compliant invoicing, WhatsApp reminders and live video classes inside the operator's own app.
Describe the system and the constraint you have hit. You will get a technical reply, not a sales sequence.
