Teams don't lose time for lack of knowledge, they lose it for lack of access.
In a thirty to eighty-person company, the critical knowledge exists: it's scattered across Notion, Drive, old threads, email. The operational cost of searching is never accounted for, but it easily exceeds the cost of a full-time hire beyond a certain headcount. The client knew this, but didn't have the number.
Upstream audit, report delivered before the engineering go-ahead.
Exhaustive mapping of knowledge sources, a user survey to quantify time lost in search, classification of query types by frequency and criticality. An eighteen-page scoping report was delivered before any development: document volumetrics, prioritised use cases, estimated annual cost of the status quo. Explicit sign-off before any code was written.
Ingestion pipeline, vector store, citation-mandatory agent.
Automated ingestion from Notion, Drive and email exports with parsing, chunking and embedding. Dedicated Qdrant vector store. Hybrid search combining semantic similarity and BM25 lexical to catch short queries. LangGraph-orchestrated agent for retrieval and synthesis, constrained to cite its sources on every answer. Docker containerisation, deployment on European infrastructure, ops documentation handed over to the internal team.