ai-contract-platform/backend/app/api/chat.py

import json
import re
from typing import Any, Dict, Iterable, List, Mapping, Optional

from fastapi import APIRouter, Depends
from fastapi.responses import StreamingResponse
from sqlalchemy.orm import Session

from app.api.reviews import require_current_user
from app.clients.shared_ai_platform import SharedAIPlatformClient, get_shared_ai_platform_client
from app.db import get_db_session
from app.schemas.chat import ChatStreamRequest
from app.services.legal_article_service import LegalArticleService
from app.services.legal_domain_registry import BUSINESS_AUXILIARY_PROMPT_KEYS
from app.services.prompt_runtime_service import PromptRuntimeService
from app.services.review_service import ReviewService
from app.services.tenant_service import TenantService

router = APIRouter(prefix="/chat", tags=["chat"])
ROUTE_PERMISSION_SCOPE = "chat"
ROUTE_PERMISSION_ACTIONS = {
    "stream": "chat:stream",
}

def get_chat_review_service(db: Session = Depends(get_db_session)) -> ReviewService:
    return ReviewService(db)


def get_chat_legal_article_service(db: Session = Depends(get_db_session)) -> LegalArticleService:
    return LegalArticleService(db)


def get_chat_prompt_runtime_service(db: Session = Depends(get_db_session)) -> PromptRuntimeService:
    return PromptRuntimeService(db)


def get_tenant_service(db: Session = Depends(get_db_session)) -> TenantService:
    return TenantService(db)


@router.post("/stream")
def stream_chat(
    payload: ChatStreamRequest,
    shared_ai_platform_client: SharedAIPlatformClient = Depends(get_shared_ai_platform_client),
    tenant_service: TenantService = Depends(get_tenant_service),
    current_user: dict = Depends(require_current_user),
    review_service: Optional[ReviewService] = Depends(get_chat_review_service),
    legal_article_service: Optional[LegalArticleService] = Depends(get_chat_legal_article_service),
    prompt_runtime_service: Optional[PromptRuntimeService] = Depends(get_chat_prompt_runtime_service),
):
    tenant_service.assert_billing_allows(current_user, "chat")
    review_context = (
        payload.review_context.model_dump(mode="json", by_alias=True, exclude_none=True)
        if payload.review_context
        else None
    )
    review_context = _enrich_review_context_with_legal_articles(current_user, review_context, review_service)
    prompt_runtime = _resolve_chat_prompt_runtime(current_user, review_context, prompt_runtime_service, payload.message)
    runtime_legal_sources = _search_runtime_legal_article_sources(
        current_user,
        payload.message,
        legal_article_service,
        prompt_runtime=prompt_runtime,
    )
    review_context = _enrich_review_context_with_runtime_legal_sources(review_context, runtime_legal_sources)
    review_context = _enrich_review_context_with_prompt_runtime(review_context, prompt_runtime, prompt_runtime_service)
    template_key = prompt_runtime.get("template_key") if prompt_runtime and prompt_runtime.get("available") else None
    if prompt_runtime_service is not None:
        prompt_runtime_service.record_run_link(
            current_user,
            prompt_runtime,
            run_type=str((prompt_runtime or {}).get("prompt_type") or "legal_chat"),
            run_id=payload.conversation_id or "new_conversation",
            resource_type="conversation",
            resource_id=payload.conversation_id,
            trace={
                "message_length": len(payload.message or ""),
                "has_review_context": bool(review_context),
                "review_id": (review_context or {}).get("reviewId") or (review_context or {}).get("review_id"),
                "template_key": template_key,
            },
        )
    local_legal_sources = _dedupe_sources([*_build_legal_article_sources(review_context), *runtime_legal_sources])
    return StreamingResponse(
        _merge_local_sources_into_stream(
            shared_ai_platform_client.stream_answer(
                current_user,
                payload.conversation_id,
                payload.message,
                review_context,
                template_key=template_key,
            ),
            local_legal_sources,
            payload.message,
        ),
        media_type="text/event-stream",
    )


def _resolve_chat_prompt_runtime(
    current_user: Dict[str, Any],
    review_context: Optional[Dict[str, Any]],
    prompt_runtime_service: Optional[PromptRuntimeService],
    message: str = "",
) -> Dict[str, Any]:
    if prompt_runtime_service is None:
        return {"available": False, "prompt_type": "legal_chat", "prompt_version": "legal_chat.shared_ai_stream.v1", "source": "local_default"}
    has_review_context = bool(review_context and ((review_context.get("reviewId") or review_context.get("review_id"))))
    business_auxiliary_key = _detect_business_auxiliary_prompt_key(message, review_context)
    if business_auxiliary_key:
        return prompt_runtime_service.resolve_published_prompt(
            current_user,
            "business_auxiliary",
            preferred_keys=[
                business_auxiliary_key,
            ],
            fallback_prompt_version="business_auxiliary.shared_ai_stream.v1",
        )
    return prompt_runtime_service.resolve_published_prompt(
        current_user,
        "legal_chat",
        preferred_keys=[
            "legal_chat_review" if has_review_context else None,
            "legal_chat_default",
        ],
        fallback_prompt_version="legal_chat.shared_ai_stream.v1",
    )


def _detect_business_auxiliary_prompt_key(
    message: str,
    review_context: Optional[Dict[str, Any]],
) -> Optional[str]:
    context_text = _build_review_context_action_text(review_context)
    if context_text:
        routed_key = _detect_business_auxiliary_prompt_key_from_text(context_text)
        if routed_key:
            return routed_key

    return _detect_business_auxiliary_prompt_key_from_text(_build_prompt_intent_text(message, review_context))


def _detect_business_auxiliary_prompt_key_from_text(text: str) -> Optional[str]:
    if _contains_any(text, ["法律意见书强边界", "法律意见书兜底", "法律意见书边界", "使用限制", "前提假设", "不能出具", "正式法律意见", "legal opinion boundary", "legal opinion fallback"]):
        return BUSINESS_AUXILIARY_PROMPT_KEYS["legal_opinion_boundary_fallback"]
    if _contains_any(text, ["履约义务", "履行义务", "义务提取", "履约清单", "义务清单", "交付义务", "配合义务", "双方要做什么", "对方需要做什么", "obligation extraction", "performance obligation", "performance checklist"]):
        return BUSINESS_AUXILIARY_PROMPT_KEYS["obligation_extraction"]
    if _contains_any(text, ["工商", "备案材料", "变更登记", "登记材料", "材料清单", "提交材料", "申报材料", "登记机关", "filing materials", "registration materials", "registration checklist", "filing checklist"]):
        return BUSINESS_AUXILIARY_PROMPT_KEYS["filing_materials_checklist"]
    if _contains_any(text, ["谈判策略", "谈判建议", "怎么谈", "如何谈", "议价", "让步方案", "让步空间", "底线", "沟通话术", "对方不接受", "替代方案", "negotiate", "negotiation strategy", "negotiation plan", "fallback position"]):
        return BUSINESS_AUXILIARY_PROMPT_KEYS["negotiation_strategy"]
    if _contains_any(text, ["法律意见书", "legal opinion"]) and _contains_any(
        text,
        ["边界", "兜底", "免责声明", "不能替代", "人工复核", "使用限制", "前提假设", "boundary", "fallback", "disclaimer", "human review", "assumption", "limitation"],
    ):
        return BUSINESS_AUXILIARY_PROMPT_KEYS["legal_opinion_boundary_fallback"]
    return None


def _build_review_context_action_text(review_context: Optional[Dict[str, Any]]) -> str:
    if not isinstance(review_context, dict):
        return ""
    metadata = review_context.get("metadata") if isinstance(review_context.get("metadata"), Mapping) else {}
    parts = [
        metadata.get("actionLabel"),
        metadata.get("focusLabel"),
        metadata.get("reviewOpinion"),
        review_context.get("actionLabel"),
        review_context.get("focusLabel"),
        review_context.get("prompt"),
    ]
    return "\n".join(str(item) for item in parts if item).lower()


def _build_prompt_intent_text(message: str, review_context: Optional[Dict[str, Any]]) -> str:
    parts = [message or ""]
    if isinstance(review_context, dict):
        parts.append(json.dumps(review_context, ensure_ascii=False, default=str))
    return "\n".join(parts).lower()


def _contains_any(value: str, candidates: Iterable[str]) -> bool:
    return any(candidate.lower() in value for candidate in candidates)


def _enrich_review_context_with_prompt_runtime(
    review_context: Optional[Dict[str, Any]],
    prompt_runtime: Optional[Dict[str, Any]],
    prompt_runtime_service: Optional[PromptRuntimeService],
) -> Optional[Dict[str, Any]]:
    if prompt_runtime_service is None or not prompt_runtime:
        return review_context
    enriched = dict(review_context or {})
    prompt_trace = prompt_runtime_service.to_trace(prompt_runtime)
    enriched["promptRuntime"] = prompt_trace
    enriched["prompt_runtime"] = prompt_trace
    metadata = dict(enriched.get("metadata") or {})
    metadata["promptRuntime"] = prompt_trace
    metadata["promptVersion"] = prompt_trace.get("prompt_version")
    metadata["promptKey"] = prompt_trace.get("prompt_key")
    enriched["metadata"] = metadata
    return enriched


def _enrich_review_context_with_legal_articles(
    current_user: Dict[str, Any],
    review_context: Optional[Dict[str, Any]],
    review_service: Optional[ReviewService],
) -> Optional[Dict[str, Any]]:
    if not review_context or review_service is None:
        return review_context

    review_id = review_context.get("reviewId") or review_context.get("review_id")
    if not review_id:
        return review_context

    try:
        result = review_service.get_review_result(current_user, str(review_id))
    except Exception:
        return review_context
    if not result:
        return review_context

    refs = _collect_legal_article_refs(result.get("risk_items") or [])
    if not refs:
        return review_context

    enriched = dict(review_context)
    enriched["legalArticleRefs"] = refs
    enriched["legal_article_refs"] = refs
    metadata = dict(enriched.get("metadata") or {})
    metadata["legalArticleRefCount"] = len(refs)
    metadata["legalArticleSources"] = [
        {
            "id": ref.get("id"),
            "title": _build_legal_article_source_title(ref),
            "detail_path": ref.get("detail_path"),
        }
        for ref in refs
    ]
    enriched["metadata"] = metadata
    return enriched


def _search_runtime_legal_article_sources(
    current_user: Dict[str, Any],
    message: str,
    legal_article_service: Optional[LegalArticleService],
    prompt_runtime: Optional[Dict[str, Any]] = None,
) -> List[Dict[str, Any]]:
    if legal_article_service is None:
        return []
    normalized_message = str(message or "").strip()
    if not normalized_message:
        return []
    if str((prompt_runtime or {}).get("prompt_type") or "").strip().lower() == "business_auxiliary":
        return []
    if not _should_search_runtime_legal_sources(normalized_message):
        return []
    try:
        return legal_article_service.search_runtime_sources(current_user, normalized_message, limit=5)
    except Exception:
        return []


def _should_search_runtime_legal_sources(message: str) -> bool:
    if re.search(r"[\u4e00-\u9fff]", message):
        return True
    lowered = message.lower()
    english_legal_markers = (
        "article",
        "law",
        "regulation",
        "civil code",
        "company law",
        "labor contract law",
        "employment law",
        "privacy law",
        "pipl",
        "data security law",
        "cybersecurity law",
        "e-signature law",
    )
    return any(marker in lowered for marker in english_legal_markers)


def _enrich_review_context_with_runtime_legal_sources(
    review_context: Optional[Dict[str, Any]],
    runtime_sources: List[Dict[str, Any]],
) -> Optional[Dict[str, Any]]:
    if not runtime_sources:
        return review_context
    enriched = dict(review_context or {})
    enriched["legalArticleSearchSources"] = runtime_sources
    enriched["legal_article_search_sources"] = runtime_sources
    metadata = dict(enriched.get("metadata") or {})
    metadata["legalArticleSearchSourceCount"] = len(runtime_sources)
    metadata["legalArticleSearchSources"] = [
        {
            "id": source.get("source_id"),
            "title": source.get("title"),
            "detail_path": source.get("detail_path"),
        }
        for source in runtime_sources
    ]
    enriched["metadata"] = metadata
    return enriched


def _collect_legal_article_refs(risk_items: Iterable[Mapping[str, Any]]) -> List[Dict[str, Any]]:
    refs: List[Dict[str, Any]] = []
    seen: set[str] = set()
    for item in risk_items:
        for ref in item.get("legal_article_refs") or []:
            if not isinstance(ref, Mapping):
                continue
            ref_id = str(ref.get("id") or "")
            key = ref_id or f"{ref.get('source_title')}::{ref.get('article_no')}::{ref.get('detail_path')}"
            if not key or key in seen:
                continue
            seen.add(key)
            refs.append(dict(ref))
    return refs


def _build_legal_article_sources(review_context: Optional[Mapping[str, Any]]) -> List[Dict[str, Any]]:
    if not review_context:
        return []
    refs = review_context.get("legalArticleRefs") or review_context.get("legal_article_refs") or []
    if not isinstance(refs, list):
        return []
    sources: List[Dict[str, Any]] = []
    for ref in refs:
        if not isinstance(ref, Mapping):
            continue
        source_id = ref.get("id") or ref.get("source_id")
        title = _build_legal_article_source_title(ref)
        content = ref.get("article_excerpt") or ref.get("match_reason") or ref.get("detail_path") or ""
        sources.append(
            {
                "title": title,
                "content": content,
                "score": 1.0,
                "source_type": "legal_article",
                "source_id": source_id,
                "detail_path": ref.get("detail_path"),
                "article_no": ref.get("article_no"),
                "authority": ref.get("authority"),
                "knowledge_base_id": ref.get("knowledge_base_id"),
                "knowledge_document_id": ref.get("knowledge_document_id"),
                "metadata": {
                    "source_id": ref.get("source_id"),
                    "source_title": ref.get("source_title"),
                    "match_reason": ref.get("match_reason"),
                    "tags": ref.get("tags") or [],
                },
            }
        )
    return _dedupe_sources(sources)


def _build_legal_article_source_title(ref: Mapping[str, Any]) -> str:
    source_title = str(ref.get("source_title") or "法条全文")
    article_no = str(ref.get("article_no") or "")
    return f"{source_title} {article_no}".strip()


def _merge_local_sources_into_stream(upstream: Iterable[str], local_sources: List[Dict[str, Any]], message: str = "") -> Iterable[str]:
    if not local_sources:
        yield from upstream
        return

    emitted_sources = False
    for chunk in upstream:
        event = _parse_sse_event(chunk)
        if event and event.get("type") == "text":
            content = str(event.get("content") or "")
            yield _format_sse_event({"type": "text", "content": _polish_chat_answer(content, message, local_sources)})
            continue
        if event and event.get("type") == "sources":
            merged = _dedupe_sources([*(event.get("sources") or []), *local_sources])
            emitted_sources = True
            yield _format_sse_event({"type": "sources", "sources": merged})
            continue
        if event and event.get("type") == "end" and not emitted_sources:
            emitted_sources = True
            yield _format_sse_event({"type": "sources", "sources": local_sources})
        yield chunk


def _polish_chat_answer(content: str, message: str, sources: List[Dict[str, Any]]) -> str:
    if not _looks_like_raw_retrieval_answer(content):
        return content
    source_titles = [str(source.get("title") or "依据来源").strip() for source in sources[:3] if source.get("title")]
    source_line = "、".join(source_titles) if source_titles else "下方引用来源"
    return (
        "我已检索到相关依据，但当前平台返回的是依据片段而不是完整法律分析。\n\n"
        f"针对你的问题“{message.strip()}”，建议先按以下方向人工核查：\n"
        "1. 明确付款节点、付款前置条件、发票要求和验收标准是否写清。\n"
        "2. 检查逾期付款、拒付、扣款和违约金规则是否具体、对等、可执行。\n"
        "3. 核对付款条款是否和交付、验收、解除、争议解决条款互相冲突。\n\n"
        f"可优先复核这些依据：{source_line}。"
    )


def _looks_like_raw_retrieval_answer(content: str) -> bool:
    normalized = content.strip().lower()
    if not normalized:
        return False
    return (
        normalized.startswith("i found relevant knowledge-base content")
        or "please treat this as ai-assisted guidance" in normalized
        or "chapter_title:" in normalized
        or "official_url:" in normalized
    )


def _parse_sse_event(chunk: str) -> Optional[Dict[str, Any]]:
    for line in chunk.splitlines():
        if not line.startswith("data: "):
            continue
        try:
            payload = json.loads(line[6:])
        except json.JSONDecodeError:
            return None
        return payload if isinstance(payload, dict) else None
    return None


def _format_sse_event(payload: Dict[str, Any]) -> str:
    return f"data: {json.dumps(payload, ensure_ascii=False)}\n\n"


def _dedupe_sources(sources: Iterable[Mapping[str, Any]]) -> List[Dict[str, Any]]:
    deduped: List[Dict[str, Any]] = []
    seen: set[str] = set()
    for source in sources:
        if not isinstance(source, Mapping):
            continue
        key = str(source.get("source_id") or source.get("detail_path") or f"{source.get('title')}::{source.get('content')}")
        if key in seen:
            continue
        seen.add(key)
        deduped.append(dict(source))
    return deduped