02Multimodal retrieval infrastructure
ZS Associates · enterprise scientific documentsScientific document knowledge layer
10,000+ concurrent users · section retrieval within 5 minutes
A retrieval layer for scientific files where the answer may live in a table, chart, formula, or a specific section of a long document. The system exposes small, composable retrieval tools for downstream agents.
HIGH LEVEL DESIGN01 — FLOW
01 · IngestScientific files1,000+ page · multimodal
02 · ParseDocument intelligenceAzure DI + custom XML handling
03 · StructureSection-aware contentDocument + section retrieval
04 · ExposeMCP retriever toolsComposable agent interface
05 · ObserveTrace each retrievalLangfuse · end-to-end
RUNTIME / SYSTEM LAYER
Amazon EKSCelery workers + message queues10,000+ concurrent users
From the constraint
to the system.
Built the document and section retriever tools over Azure Document Intelligence and custom XML handling; instrumented the workflow with Langfuse and deployed the retrieval workers on EKS with Celery and message queues.
END-TO-END PATH
01Multimodal scientific file
02Extract text, structure and layout
03Index document and section context
04Retrieve through MCP tools
05Return sourced context to agent