RAG Knowledge Agent with Persistent Memory

Built a Retrieval-Augmented Generation knowledge agent that ingests organisational documents (reports, exported email threads, meeting notes), builds a semantic vector store, and answers complex cross-document queries with inline source citations and confidence indicators.

Replaced naive fixed-length chunking with a structure-aware heuristic that never splits inside a sentence and never separates a table from its header row, then exposed the retrieved chunks directly in the UI so operators can distinguish retrieval bugs from generation bugs in seconds.

Session memory persists across conversations via a per-session vector namespace, retrieved only when the new query references prior turns, which stopped the agent from leaking stale context into unrelated questions. An incremental ingestion pipeline keyed on content hashes processes only changed files and writes the manifest atomically for crash safety.