Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/rag/service.py

@@ -58,6 +58,59 @@ class RetrievalService:
         self._section_resolver = section_resolver
         self._reranker = reranker
 
+    def retrieve_framed(
+        self,
+        drug_id: str,
+        section_key: str | None,
+        query: str,
+        is_overview: bool = False,
+    ) -> RetrievalResult:
+        """Retrieve driven by an already-understood frame, not by parsing text.
+
+        The LLM understanding layer has resolved the drug (against the real
+        catalog) and, when the turn named one, the section. So this skips the
+        fuzzy `CatalogDrugResolver` and the keyword `SectionResolver` entirely:
+        a named `section_key` filters that section whole; without one, this
+        mirrors `retrieve()`'s two remaining cases — `is_overview` (the frame's
+        `turn_type == "drug_overview"`, a bare name) answers from the identity
+        sections only, and a free-form question reranks the full monograph
+        down to `rerank_top_k`.
+
+        Found live 2026-08-06: without the `is_overview` split, a bare drug
+        name sent the ENTIRE ~29-section monograph as evidence for every
+        generation call (retrieval had no notion of "just the intro"), which
+        is both wrong retrieval and, downstream, an answer so long it
+        intermittently failed generation/entailment outright. `query` is only
+        the rerank signal here, never a resolution input.
+        """
+        if not drug_id.strip():
+            return RetrievalResult(EvidenceDecision.ABSTAIN, "missing_query_or_drug")
+
+        if section_key:
+            find_by_section = getattr(self._retriever, "find_by_section", None)
+            if find_by_section is not None:
+                hits = find_by_section(drug_id, section_key)
+                if hits:
+                    return self._decide(self._hydrate(hits, limit=None))
+
+        overview_hits = self._drug_overview(drug_id)
+        if overview_hits is None:
+            return RetrievalResult(EvidenceDecision.ABSTAIN, "insufficient_retrieval_score")
+
+        if is_overview:
+            intro = [h for h in overview_hits if h.document.section_key in INTRO_SECTIONS]
+            return self._decide(
+                self._hydrate(intro or overview_hits, limit=None), is_drug_overview=True
+            )
+
+        overview_hits = self._rerank(query, overview_hits)
+        # Capped even when rerank is disabled/unavailable and fails open to
+        # the unfiltered list — an ordering aid must never remove the size
+        # bound too, or the same 29-section explosion returns through here.
+        return self._decide(
+            self._hydrate(overview_hits, limit=self._policy.evidence_limit)
+        )
+
     def retrieve(self, query: str, drug_id: str) -> RetrievalResult:
         if not query.strip() or not drug_id.strip():
             return RetrievalResult(EvidenceDecision.ABSTAIN, "missing_query_or_drug")