Retrieval-augmented generation works beautifully on document corpora where topics are distinct: a legal brief does not read like a product spec. Organizational communications are the opposite. Every channel, from a sales thread to a standup recap, is written in the same register, about the same people, using the same shorthand.
When the surface language is homogeneous, semantic similarity stops discriminating. The nearest neighbors to "which deals are stuck?" are a hundred messages that all sound like deals, none of which answer the question. Cosine distance measures how things are said, not what is true.
Signal takes a different path. At ingest we normalize communications into structured fields, then answer questions by tool-calling against that structure rather than embedding-matching against raw text. The retrieval problem becomes a query problem, and query problems have exact answers.