Spaces:
Sleeping
Sleeping
| from __future__ import annotations | |
| from typing import List, Optional | |
| from pydantic import BaseModel, Field | |
| from app.agents.cerebras_client import CerebrasClient | |
| from app.schemas.graph import FormulaContent, FormulaStep, HTML5VisualPayload | |
| _MAX_CHUNK_CHARS = 3000 # mirrors ModalityRouter._MAX_CHUNK_CHARS / BrainAgent.extract_curriculum cap | |
| _DEFAULT_DECLINE_REASON = "This concept doesn't have an explicit formula in the source — better explored in chat." | |
| class FormulaGrounding(BaseModel): | |
| renderable: bool | |
| anchor: str = "" | |
| main_latex: str = "" | |
| steps: List[FormulaStep] = Field(default_factory=list) | |
| decline_reason: str = "" | |
| class FormulaEngine: | |
| """Two-stage grounded formula extraction: extract (with a self-checked verbatim | |
| anchor) then render into an HTML5VisualPayload. Mirrors the D3TemplateRouter/ | |
| D3DataExtractor split and the old TutorAgent visual-grounding pipeline.""" | |
| def __init__(self, client: Optional[CerebrasClient] = None) -> None: | |
| self._client = client or CerebrasClient() | |
| def generate(self, concept: str, chunks: List[dict], familiarity: str) -> HTML5VisualPayload: | |
| from app.agents.tutor_agent import TutorAgent | |
| grounding = self._extract(concept, chunks, familiarity) | |
| if not grounding.renderable: | |
| reason = grounding.decline_reason or _DEFAULT_DECLINE_REASON | |
| return HTML5VisualPayload( | |
| html_code=TutorAgent._decline_html(concept, reason), | |
| animation_type="2d_text", | |
| explanation=reason, | |
| ) | |
| return HTML5VisualPayload( | |
| html_code="", | |
| animation_type="formula", | |
| formula=FormulaContent(main_latex=grounding.main_latex, steps=grounding.steps), | |
| explanation=(f"From the source: {grounding.anchor}" if grounding.anchor else ""), | |
| source="paper" if grounding.anchor else "model_knowledge", | |
| ) | |
| def _extract(self, concept: str, chunks: List[dict], familiarity: str) -> FormulaGrounding: | |
| chunk_text = "\n\n".join(f"[Source: {c.get('source', '?')}]: {c['text']}" for c in chunks)[:_MAX_CHUNK_CHARS] | |
| messages = [ | |
| {"role": "system", "content": ( | |
| "You extract a grounded formula/equation for a concept from the SOURCE MATERIAL only. " | |
| "If the source contains an explicit formula/equation for this concept, copy a VERBATIM " | |
| "30-80 character substring of the source's actual equation/formula line as `anchor` — " | |
| "it must match the source text literally, not a paraphrase. Fill `main_latex` and `steps` " | |
| "using ONLY numbers, coefficients, and notation actually present in the source — never " | |
| "invent or estimate values. If no explicit formula/equation exists in the source, set " | |
| "renderable=false and give a short, calm, student-facing `decline_reason` (e.g. " | |
| f"\"'{concept}' doesn't have an explicit formula in the source — better explored in " | |
| "chat.\").\n\n" | |
| f"SOURCE MATERIAL:\n{chunk_text}" | |
| )}, | |
| {"role": "user", "content": f"Concept: '{concept}' (level: {familiarity}). Extract the formula or decline."}, | |
| ] | |
| grounding = self._client.structured_complete(messages, FormulaGrounding, reasoning_effort="medium") | |
| if grounding.anchor and grounding.anchor.strip() not in chunk_text: | |
| grounding = grounding.model_copy(update={"anchor": ""}) | |
| return grounding | |