Filter a list of sentences to only those relevant to query.

Use this BEFORE feeding noisy context (earnings-call transcripts, news articles, long reports, log lines) to an LLM; it typically cuts token spend without losing important context. Chain it with your own language model.

Args: sentences: raw input sentences; ~2000 per call is comfortable. query: natural-language description of the topic to filter for (e.g. "AI capex plans", "rate cuts", "supply chain risk").

Returns: dict with: relevant_sentences: sentences that passed the filter, original order. total_sentences, relevant_count: counts to compute savings. topic_used: the topic string the model actually filtered against. topic_source: "zettaquant" (server-derived) or "caller" (you supplied). model_id: which V-SLM head handled the request.