MIYA Chat Intelligence

Overview

MIYA Chat Intelligence connects conversational AI with the controls needed to manage enterprise knowledge.

Users can ask questions over published documents, explore the knowledge directory, and analyze uploaded files with persistent conversation history. Administrators manage document lifecycles, review unanswered questions and feedback, and monitor usage, latency, and token consumption. A Next.js dashboard and FastAPI single-agent backend integrate with Azure services for document processing, retrieval, and persistence. Conversational access to governed enterprise knowledge Business context Enterprise knowledge tools serve employees seeking answers and administrators responsible for content quality. MIYA brings document availability, chat history, feedback, and operational monitoring into a shared dashboard. Objectives - Enable questions over published knowledge and uploaded files. - Keep retrieval aligned with publication, hierarchy, and content versions. - Provide visibility into unanswered questions, feedback, and usage. Solution architecture - Next.js workspace Authenticated React dashboard and session-aware browser BFF. - FastAPI and chat agent Knowledge write authority and a single agent with document, directory, uploaded-file, and unanswered-question tools. - Document processing Blob Storage, Document Intelligence, embeddings, and optional Durable Functions support the indexing lifecycle. - Knowledge retrieval Azure AI Search supplies document chunks subject to publication, version, scope, and projection-readiness checks. - Application persistence Cosmos DB stores application records, knowledge metadata, operations, and audit records. Workflow - Organize knowledge Manage folders, files, and publication in the content library. - Process and index Extract document content, generate embeddings, and synchronize the search projection. - Ask and retrieve Search published knowledge or uploaded files, with document-specific or recursive folder scope when needed. - Review and improve Administrators review unanswered questions, feedback, conversation logs, and usage analytics. Engineering challenges - Prevent stale retrieval Document changes and asynchronous indexing can leave search temporarily out of step with authoritative metadata. Fail closed on invalid node state, content version, scope revision, or search-projection readiness. Content can remain unavailable during synchronization rather than expose stale results. - Separate service responsibilities Browser authentication, knowledge writes, and long-running processing require distinct boundaries. Use Next.js for browser sessions, FastAPI for knowledge authority, and optional Durable Functions for lifecycle orchestration. Multiple services require coordinated configuration and monitoring. Lessons learned - Retrieval correctness depends on lifecycle state as well as relevance. - Human review of unanswered questions complements automated retrieval. - Separate implemented capabilities from unverified business outcomes.