Spaces:
Sleeping
Sleeping
| """ContextPack assembler. | |
| Composes deterministic packs from retrieval results: | |
| - course_constitution: all course_meta records (always loaded) | |
| - topic: retrieval top-K | |
| - module: full topic content (intent-driven) | |
| assessment_safe is enforced upstream (catalog only contains student_access=True | |
| records by construction in sub-project #1's _student_visible_files). | |
| """ | |
| from __future__ import annotations | |
| from dataclasses import dataclass, field | |
| from scripts.lib.catalog_loader import Catalog | |
| from scripts.lib.ranker import RetrievalResult | |
| _ASSESSMENT_KINDS = {"question", "slide_answer"} | |
| _MODULE_INTENTS = {"summarize", "explain"} | |
| class ContextPack: | |
| course_constitution: list[dict] = field(default_factory=list) | |
| topic: list[RetrievalResult] = field(default_factory=list) | |
| module: list[dict] = field(default_factory=list) | |
| has_assessment_content: bool = False | |
| has_redacted_tier: bool = False | |
| def assemble( | |
| *, | |
| intent: str, | |
| retrieved: list[RetrievalResult], | |
| catalog: Catalog, | |
| mode: str = "tutor", | |
| ) -> ContextPack: | |
| """Compose a ContextPack from an intent + retrieval results + mode. | |
| In review mode the module block is expanded whenever the dominant-topic | |
| threshold is met, regardless of intent — faculty queries about a concept | |
| consistently want every occurrence across the term, not the intent-driven | |
| subset that's right for student tutoring. | |
| """ | |
| by_doc_id = {r["doc_id"]: r for r in catalog.records} | |
| constitution = [r for r in catalog.records if r.get("kind") == "course_meta"] | |
| topic = list(retrieved) | |
| module: list[dict] = [] | |
| module_eligible = mode == "review" or intent in _MODULE_INTENTS | |
| if module_eligible and len(retrieved) >= 2: | |
| # Find the topic with the most retrieved records | |
| topic_counts: dict[str, int] = {} | |
| for r in retrieved: | |
| rec = by_doc_id.get(r.doc_id, {}) | |
| t = rec.get("topic", "") | |
| if t: | |
| topic_counts[t] = topic_counts.get(t, 0) + 1 | |
| if topic_counts: | |
| dominant_topic = max(topic_counts, key=lambda k: topic_counts[k]) | |
| if topic_counts[dominant_topic] >= 2: | |
| module = [ | |
| rec for rec in catalog.records | |
| if rec.get("topic") == dominant_topic | |
| ] | |
| has_assessment = any( | |
| by_doc_id.get(r.doc_id, {}).get("kind") in _ASSESSMENT_KINDS | |
| for r in retrieved | |
| ) | |
| has_redacted = any( | |
| by_doc_id.get(r.doc_id, {}).get("privacy_tier") == "redacted_classroom_speech" | |
| for r in retrieved | |
| ) | |
| return ContextPack( | |
| course_constitution=constitution, | |
| topic=topic, | |
| module=module, | |
| has_assessment_content=has_assessment, | |
| has_redacted_tier=has_redacted, | |
| ) | |