feat(bot): tool-using chat agent over RAG, ЕМИАС record and drug-safety reference
CI / ruff · pytest (pull_request) Successful in 2m10s

Free-text treatment and medication questions go to an LLM agent with read-only tools bound to the current
patient: guideline search (RAG, doses stripped), ЕМИАС record mock (diagnoses, comorbidities, intolerances,
other specialists' drugs, current prescription), drug info (composition, contraindications, interactions)
and a deterministic patient-specific check. Providers: OpenRouter (temporary, synthetic data only) and an
offline mock; GigaChat to follow. Scrubbed input only, forbidden-output filter, any failure falls back to
the existing extractive RAG / instruction answers. Off unless OPENROUTER_API_KEY or CHAT_AGENT=mock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-10-04 14:39:16 +03:00
co-authored by Claude Opus 5.5
parent 2f1f8a8f24
commit 6dda5f9b30
14 changed files with 1253 additions and 2 deletions
+179
View File
@@ -0,0 +1,179 @@
"""Read-only tools of the chat agent. Every tool is bound to the current patient's token on the server side:
the model never passes a token, so it cannot read another patient's data (medical secrecy, 323-FZ art. 13).
Tool results contain clinical data only (no name, SNILS, phone) — they are sent to the LLM provider.
"""
import json
import logging
import re
from dataclasses import dataclass, field
from typing import Any
from app.integrations.emias_record import PatientRecord, emias_records
from app.llm.agent.drug_safety import check_patient, find_drug, load_db
from app.llm.rag import retrieve
from app.llm.rag.index import sentences
log = logging.getLogger(__name__)
RAG_TOP_K = 4
PASSAGE_CHARS = 700
# Patients never get dosages (SPEC §7): sentences with doses are dropped before the text reaches the LLM.
_DOSE = re.compile(r"\d\s*(?:-|–)?\s*\d*\s*(?:мг|mg|мкг|г/сут|таблетк)|\bдоз(?:а|у|ы|е|ой|ировк)", re.I)
TOOL_SPECS: list[dict[str, Any]] = [
{
"type": "function",
"function": {
"name": "search_guidelines",
"description": (
"Поиск по базе знаний: клинические рекомендации Минздрава РФ (депрессия 2024, ПТСР 2023). "
"Возвращает фрагменты с источником и страницей. Используй для любых общих вопросов о состоянии, "
"лечении, сроках действия терапии, побочных эффектах, отмене, образе жизни."
),
"parameters": {
"type": "object",
"properties": {"query": {"type": "string", "description": "Поисковый запрос на русском"}},
"required": ["query"],
},
},
},
{
"type": "function",
"function": {
"name": "get_patient_record",
"description": (
"Медицинская карта текущего пациента из ЕМИАС: диагнозы (МКБ-10), сопутствующие заболевания, "
"аллергии и непереносимости, назначения других врачей и текущее назначение психиатра. "
"Вызывай, когда ответ зависит от особенностей пациента или его лекарств."
),
"parameters": {"type": "object", "properties": {}},
},
},
{
"type": "function",
"function": {
"name": "get_drug_info",
"description": (
"Справка о препарате по инструкции ГРЛС: состав (действующее и вспомогательные вещества), "
"противопоказания, взаимодействия, беременность. Аргумент — название препарата (МНН или торговое)."
),
"parameters": {
"type": "object",
"properties": {"drug": {"type": "string"}},
"required": ["drug"],
},
},
},
{
"type": "function",
"function": {
"name": "check_drug_for_patient",
"description": (
"Детерминированная проверка препарата для текущего пациента: сопутствующие заболевания, "
"непереносимости по составу, взаимодействия с другими его лекарствами. Используй перед любым "
"ответом о том, подходит ли пациенту препарат или сочетание препаратов."
),
"parameters": {
"type": "object",
"properties": {"drug": {"type": "string"}},
"required": ["drug"],
},
},
},
]
@dataclass
class ToolContext:
token: str
sources: list[str] = field(default_factory=list) # citations collected for the answer footer
used: list[str] = field(default_factory=list)
_record: PatientRecord | None = None
_record_loaded: bool = False
async def record(self) -> PatientRecord | None:
if not self._record_loaded:
self._record = await emias_records.get(self.token)
self._record_loaded = True
return self._record
def cite(self, label: str) -> None:
if label and label not in self.sources:
self.sources.append(label)
async def run_tool(ctx: ToolContext, name: str, args: dict[str, Any]) -> str:
"""Execute a tool; always returns a JSON string (errors included) so the model can recover."""
ctx.used.append(name)
log.info("agent tool=%s", name) # no arguments: they may echo patient text
try:
if name == "search_guidelines":
return _json(_search(ctx, str(args.get("query", ""))))
if name == "get_patient_record":
rec = await ctx.record()
if rec is None:
return _json({"found": False})
ctx.cite("ЕМИАС: медицинская карта")
return _json({"found": True, **rec.model_dump()})
if name == "get_drug_info":
return _json(_drug_info(ctx, str(args.get("drug", ""))))
if name == "check_drug_for_patient":
return _json(await _check(ctx, str(args.get("drug", ""))))
return _json({"error": f"unknown tool {name}"})
except Exception as exc: # noqa: BLE001 — a tool failure must not crash the chat
log.error("agent tool=%s failed: %s", name, type(exc).__name__)
return _json({"error": "tool failed"})
def _search(ctx: ToolContext, query: str) -> dict[str, Any]:
hits = retrieve(query, audience="patient", top_k=RAG_TOP_K) if query.strip() else []
out = []
for h in hits:
label = h.chunk.citation.strip("[]")
ctx.cite(label)
text = " ".join(x for x in sentences(h.passage) if not _DOSE.search(x))[:PASSAGE_CHARS]
if text:
out.append({"source": label, "section": h.chunk.section, "text": text})
return {"results": out, "found": bool(out)}
def _drug_info(ctx: ToolContext, name: str) -> dict[str, Any]:
entry = find_drug(name)
if entry is None:
return {"found": False, "drug": name}
ctx.cite(entry.source)
return {
"found": True,
"inn": entry.inn,
"class": entry.drug_class,
"composition": entry.composition.model_dump() if entry.composition else None,
"contraindications": entry.contraindications,
"interactions": [i.text for i in entry.interactions],
"pregnancy": entry.pregnancy,
"source": entry.source,
}
async def _check(ctx: ToolContext, name: str) -> dict[str, Any]:
entry = find_drug(name)
if entry is None:
return {"found": False, "drug": name, "note": "Препарата нет в справочнике — отвечать «уточните у врача»."}
rec = await ctx.record()
if rec is None:
return {"found": True, "inn": entry.inn, "record": False, "findings": []}
ctx.cite(entry.source)
ctx.cite("ЕМИАС: медицинская карта")
findings = check_patient(entry, rec, load_db())
return {
"found": True,
"inn": entry.inn,
"record": True,
"findings": [f.model_dump() for f in findings],
"note": "Пустой список не означает, что препарат безопасен: решение о приёме принимает врач.",
}
def _json(data: Any) -> str:
return json.dumps(data, ensure_ascii=False)