feat(api): 统一 LLM JSON 调用层 llm_json_call，按域 Schema 迁移 chat/memoir agents

2026-04-03 13:34:27 +08:00
parent 41518bda11
commit 43d1689e9c
28 changed files with 1006 additions and 352 deletions
--- a/api/app/agents/chat/profile_agent.py
+++ b/api/app/agents/chat/profile_agent.py
@@ -3,7 +3,6 @@ ProfileAgent：用户资料收集 Specialist
 负责提取资料、资料追问、资料收集开场白，不负责 Redis 持久化（由 Orchestrator 统一处理）
 """

-import json
 import time
 from typing import Any, Dict, List, Optional

@@ -15,11 +14,11 @@ from app.agents.chat.prompts_profile import (
    get_profile_followup_prompt,
    get_profile_greeting_prompt,
 )
+from app.agents.chat.schemas import ProfileExtractionOutput
 from app.core.agent_logging import agent_span, log_agent_payload, log_agent_summary
 from app.core.config import settings
 from app.core.dependencies import get_llm_provider
-from app.core.json_utils import extract_json_payload
-from app.core.langchain_llm import ainvoke_json_object
+from app.core.llm_call import allm_json_call
 from app.core.logging import get_logger
 from app.agents.chat.reply_limits import (
    nonempty_segments_or_fallback,
@@ -53,6 +52,53 @@ class ProfileAgent:
    def __init__(self):
        self.llm = _get_langchain_llm()

+    async def _invoke_chat(
+        self,
+        messages: List[Any],
+        *,
+        max_tokens: int,
+        conversation_id: Optional[str],
+        agent_name: str,
+    ) -> str:
+        chat_llm = self.llm.bind(max_tokens=max_tokens)
+        llm_t0 = time.perf_counter()
+        with agent_span(
+            logger, f"{agent_name}.llm", conversation_id=conversation_id or ""
+        ):
+            response = await chat_llm.ainvoke(messages)
+        logger.info(
+            "event=chat_llm_done agent={} response_latency_ms={:.2f}",
+            agent_name,
+            (time.perf_counter() - llm_t0) * 1000,
+        )
+        return (
+            response.content if hasattr(response, "content") else str(response)
+        ) or ""
+
+    async def _segments_from_response(
+        self,
+        response_text: str,
+        *,
+        max_segments: int,
+        max_chars_per_segment: int,
+        fallback: str,
+    ) -> List[str]:
+        log_agent_payload(
+            logger,
+            "ProfileAgent._segments_from_response.raw_response",
+            response_text,
+        )
+        raw_list = segments_from_llm_response(response_text, max_segments=max_segments)
+        if not raw_list:
+            raw_list = [response_text.strip()]
+        out = truncate_chat_segments(
+            raw_list,
+            max_segments=max_segments,
+            max_chars_per_segment=max_chars_per_segment,
+        )
+        segments = out if out else [response_text.strip()[:max_chars_per_segment]]
+        return nonempty_segments_or_fallback(segments, fallback=fallback)
+
    async def extract_profile_from_message(
        self,
        user_message: str,
@@ -81,16 +127,17 @@ class ProfileAgent:
            prompt = get_profile_extraction_prompt(
                user_message, missing_fields, recent_dialogue=recent_dialogue or None
            )
-            content = await ainvoke_json_object(
+            parsed = await allm_json_call(
                self.llm,
                prompt,
-                max_tokens=512,
+                ProfileExtractionOutput,
+                max_tokens=settings.chat_profile_extract_max_tokens,
                agent="ProfileAgent.extract_profile_from_message",
+                fallback_factory=lambda: ProfileExtractionOutput(),
            )
-            parsed = json.loads(extract_json_payload(content))
            result = {}
-            if "birth_year" in parsed and parsed["birth_year"] is not None:
-                raw = parsed["birth_year"]
+            if parsed.birth_year is not None:
+                raw = parsed.birth_year
                if isinstance(raw, int) and 1900 <= raw <= 2100:
                    result["birth_year"] = raw
                elif isinstance(raw, str) and raw.isdigit():
@@ -99,14 +146,14 @@ class ProfileAgent:
                        y = 1900 + y if y >= 50 else 2000 + y
                    if 1900 <= y <= 2100:
                        result["birth_year"] = y
-            if "birth_place" in parsed and parsed["birth_place"]:
-                result["birth_place"] = str(parsed["birth_place"])
-            if "grew_up_place" in parsed and parsed["grew_up_place"]:
-                result["grew_up_place"] = str(parsed["grew_up_place"])
-            if "occupation" in parsed and parsed["occupation"]:
-                result["occupation"] = str(parsed["occupation"])
+            if parsed.birth_place:
+                result["birth_place"] = str(parsed.birth_place)
+            if parsed.grew_up_place:
+                result["grew_up_place"] = str(parsed.grew_up_place)
+            if parsed.occupation:
+                result["occupation"] = str(parsed.occupation)
            return result
-        except (json.JSONDecodeError, Exception) as e:
+        except Exception as e:
            logger.error("提取资料信息失败: {}", e)
            return {}

@@ -143,61 +190,33 @@ class ProfileAgent:
                "ProfileAgent.followup.prompt",
                format_history_string(messages),
            )
-            chat_llm = self.llm.bind(
-                max_tokens=settings.chat_profile_followup_max_tokens
-            )
-            llm_t0 = time.perf_counter()
-            with agent_span(
-                logger,
-                "ProfileAgent.followup.llm",
-                conversation_id=conversation_id,
-            ):
-                logger.info(
-                    "event=chat_prompt_built agent=ProfileAgent.generate_profile_followup "
-                    "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
-                    _message_contents_char_count(messages),
-                    hw.turn_total,
-                    len(hw.window) // 2,
-                )
-                response = await chat_llm.ainvoke(messages)
+            prompt_chars = _message_contents_char_count(messages)
            logger.info(
-                "event=chat_llm_done agent=ProfileAgent.generate_profile_followup "
-                "response_latency_ms={:.2f}",
-                (time.perf_counter() - llm_t0) * 1000,
+                "event=chat_prompt_built agent=ProfileAgent.generate_profile_followup "
+                "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
+                prompt_chars,
+                hw.turn_total,
+                len(hw.window) // 2,
            )
-            response_text = (
-                response.content if hasattr(response, "content") else str(response)
+            response_text = await self._invoke_chat(
+                messages,
+                max_tokens=settings.chat_profile_followup_max_tokens,
+                conversation_id=conversation_id,
+                agent_name="ProfileAgent.generate_profile_followup",
            )
-            log_agent_payload(
-                logger, "ProfileAgent.followup.raw_response", response_text
-            )
-            raw_list = segments_from_llm_response(response_text, max_segments=3)
-            if not raw_list:
-                raw_list = [response_text.strip()]
-            out = truncate_chat_segments(
-                raw_list,
+            segments = await self._segments_from_response(
+                response_text,
                max_segments=3,
                max_chars_per_segment=settings.chat_interview_max_chars_per_segment,
+                fallback="谢谢分享！能再告诉我一些吗？",
            )
            log_agent_summary(
                logger,
                "ProfileAgent.followup segments={} conversation_id={}",
-                len(out),
+                len(segments),
                conversation_id,
            )
-            segments = (
-                out
-                if out
-                else [
-                    response_text.strip()[
-                        : settings.chat_interview_max_chars_per_segment
-                    ]
-                ]
-            )
-            return nonempty_segments_or_fallback(
-                segments,
-                fallback="谢谢分享！能再告诉我一些吗？",
-            )
+            return segments
        except Exception as e:
            logger.error("生成资料跟进回复失败: {}", e)
            return ["谢谢分享！能再告诉我一些吗？"]
@@ -229,61 +248,33 @@ class ProfileAgent:
            log_agent_payload(
                logger, "ProfileAgent.greeting.prompt", format_history_string(messages)
            )
-            chat_llm = self.llm.bind(
-                max_tokens=settings.chat_profile_followup_max_tokens
-            )
-            llm_t0 = time.perf_counter()
-            with agent_span(
-                logger,
-                "ProfileAgent.greeting.llm",
-                conversation_id=conversation_id,
-            ):
-                logger.info(
-                    "event=chat_prompt_built agent=ProfileAgent.generate_profile_greeting "
-                    "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
-                    _message_contents_char_count(messages),
-                    hw.turn_total,
-                    len(hw.window) // 2,
-                )
-                response = await chat_llm.ainvoke(messages)
+            prompt_chars = _message_contents_char_count(messages)
            logger.info(
-                "event=chat_llm_done agent=ProfileAgent.generate_profile_greeting "
-                "response_latency_ms={:.2f}",
-                (time.perf_counter() - llm_t0) * 1000,
+                "event=chat_prompt_built agent=ProfileAgent.generate_profile_greeting "
+                "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
+                prompt_chars,
+                hw.turn_total,
+                len(hw.window) // 2,
            )
-            response_text = (
-                response.content if hasattr(response, "content") else str(response)
+            response_text = await self._invoke_chat(
+                messages,
+                max_tokens=settings.chat_profile_followup_max_tokens,
+                conversation_id=conversation_id,
+                agent_name="ProfileAgent.generate_profile_greeting",
            )
-            log_agent_payload(
-                logger, "ProfileAgent.greeting.raw_response", response_text
-            )
-            raw_list = segments_from_llm_response(response_text, max_segments=2)
-            if not raw_list:
-                raw_list = [response_text.strip()]
-            out = truncate_chat_segments(
-                raw_list,
+            segments = await self._segments_from_response(
+                response_text,
                max_segments=2,
                max_chars_per_segment=settings.chat_interview_max_chars_per_segment,
+                fallback="你好！在开始之前，能告诉我你是哪一年出生的吗？",
            )
            log_agent_summary(
                logger,
                "ProfileAgent.greeting segments={} conversation_id={}",
-                len(out),
+                len(segments),
                conversation_id,
            )
-            segments = (
-                out
-                if out
-                else [
-                    response_text.strip()[
-                        : settings.chat_interview_max_chars_per_segment
-                    ]
-                ]
-            )
-            return nonempty_segments_or_fallback(
-                segments,
-                fallback="你好！在开始之前，能告诉我你是哪一年出生的吗？",
-            )
+            return segments
        except Exception as e:
            logger.error("生成资料收集开场白失败: {}", e)
            return [