Download src/agent.py from sarra2002/RareDx: direct link, hf CLI and curl.
- Browser
- Download file 3.88 kB
-
https://huggingface.co/spaces/sarra2002/RareDx/resolve/main/src/agent.py
- Command line
-
hf download hf://spaces/sarra2002/RareDx/src/agent.py
-
curl -L -o agent.py https://huggingface.co/spaces/sarra2002/RareDx/resolve/main/src/agent.py
3.88 kB
| """Agent v2: ask the doctor for the findings or lab results that best separate the top hypotheses.""" | |
| from __future__ import annotations | |
| import numpy as np | |
| from .system import DiagnosticSystem | |
| class QuestionAgent: | |
| def __init__(self, system: DiagnosticSystem, n_questions: int = 5, head: int = 10, window: int = 20, | |
| confirm_weight: float = 1.0, exclude_weight: float = 0.5, lab_bonus: float = 1.5, | |
| min_ic: float = 2.0): | |
| self.system, self.kb, self.rag = system, system.kb, system.rag | |
| self.n_questions, self.head, self.window = n_questions, head, window | |
| self.confirm_weight, self.exclude_weight = confirm_weight, exclude_weight # λ and μ | |
| self.lab_bonus, self.min_ic = lab_bonus, min_ic | |
| self.choose_questions = system.tracer.wrap(self.choose_questions, "agent : choix des questions") | |
| self.update = system.tracer.wrap(self.update, "agent : mise à jour après réponses") | |
| def choose_questions(self, ranking: list[str], codes: list[str]) -> list[str]: | |
| """Findings with the largest rank-weighted variance across the top diseases (labs get a bonus).""" | |
| top = ranking[:self.head] | |
| if not top: | |
| return [] | |
| p = np.array([1 / (1 + r) for r in range(1, len(top) + 1)]) | |
| p /= p.sum() | |
| weights = [self.kb.extended_weights(g) for g in top] | |
| known = self.kb.expand(codes) | |
| notes = {} | |
| for t in set().union(*weights) - known: | |
| if self.rag.ic.get(t, 0) < self.min_ic: # too general | |
| continue | |
| v = np.array([w.get(t, 0) for w in weights]) | |
| note = (p * (v - (p * v).sum()) ** 2).sum() | |
| notes[t] = note * (self.lab_bonus if self.kb.is_metabolic(t) else 1) | |
| chosen: list[str] = [] | |
| for t in sorted(notes, key=lambda t: (-notes[t], t)): # ties broken by code: reproducible | |
| if all(t not in self.kb.ancestors(c) and c not in self.kb.ancestors(t) for c in chosen): | |
| chosen.append(t) | |
| if len(chosen) == self.n_questions: | |
| break | |
| return chosen | |
| def rerank(self, ranking: list[str], confirmed: list[str], excluded: list[str]) -> list[str]: | |
| scored = [] | |
| for r, g in enumerate(ranking[:self.window], 1): | |
| w, s = self.kb.extended_weights(g), 1 / (1 + r) | |
| for c in confirmed: | |
| s *= 1 + self.confirm_weight * w.get(c, 0) | |
| for c in excluded: | |
| s *= 1 - self.exclude_weight * w.get(c, 0) | |
| scored.append((s, g)) | |
| scored.sort(key=lambda x: -x[0]) | |
| return [g for _, g in scored] + ranking[self.window:] | |
| def update(self, codes: list[str], llm_sets: list[set[str]], confirmed: list[str], excluded: list[str]) -> list[str]: | |
| """Add the confirmed findings, re-run the RAG, fuse again (no new LLM call), then rerank.""" | |
| codes2 = list(dict.fromkeys(codes + confirmed)) | |
| ranking = self.system.fuse(self.system.search_rag(codes2), llm_sets, codes2) | |
| return self.rerank(ranking, confirmed, excluded) | |
| def read_answers(self, questions: list[str], answers: dict[int, str]) -> tuple[list[str], list[str]]: | |
| """Doctor's answers: "non" / "no", "?" (unknown), or a precise result in free text.""" | |
| confirmed, excluded = [], [] | |
| for i, q in enumerate(questions, 1): | |
| a = str(answers.get(i, "?")).strip() | |
| if a.lower() in ("?", "je ne sais pas", "unknown", ""): | |
| continue | |
| if a.lower() in ("non", "no"): | |
| excluded.append(q) | |
| continue | |
| precise = [r[0] for p in self.system.llm.extract_symptoms(a) if (r := self.system.linker.to_hpo(p))] | |
| confirmed += [c for c in (precise or [q]) if c not in confirmed] | |
| return confirmed, excluded | |