feat: 统一提示词模板并添加全量提示词日志 [AC-AISVC-02, AC-ASA-19, AC-ASA-20]
PR Check (SDD Full Gate) / sdd-full-gate (pull_request) Successful in 2s
Details
PR Check (SDD Full Gate) / sdd-full-gate (pull_request) Successful in 2s
Details
- 新增 prompts.py 集中管理系统提示词和证据格式化 - orchestrator.py 添加全量提示词日志打印 - openai_client.py 添加全量提示词日志打印（支持普通和流式） - rag.py 重构使用统一的提示词模板
This commit is contained in:
MerCry
parent
dd74ae2585
commit
f631f1dea0
|
|
@ -14,6 +14,7 @@ from pydantic import BaseModel, Field
|
|||
|
||||
from app.core.config import get_settings
|
||||
from app.core.exceptions import MissingTenantIdException
|
||||
from app.core.prompts import format_evidence_for_prompt, build_user_prompt_with_evidence
|
||||
from app.core.tenant import get_tenant_id
|
||||
from app.models import ErrorResponse
|
||||
from app.services.retrieval.vector_retriever import get_vector_retriever
|
||||
|
|
@ -226,6 +227,11 @@ async def run_rag_experiment_stream(
|
|||
|
||||
final_prompt = _build_final_prompt(request.query, retrieval_results)
|
||||
|
||||
logger.info(f"[AC-ASA-20] ========== RAG LAB STREAM FULL PROMPT ==========")
|
||||
logger.info(f"[AC-ASA-20] Prompt length: {len(final_prompt)}")
|
||||
logger.info(f"[AC-ASA-20] Prompt content:\n{final_prompt}")
|
||||
logger.info(f"[AC-ASA-20] ==============================================")
|
||||
|
||||
yield f"event: retrieval\ndata: {json.dumps({'results': retrieval_results, 'count': len(retrieval_results)})}\n\n"
|
||||
|
||||
yield f"event: prompt\ndata: {json.dumps({'prompt': final_prompt})}\n\n"
|
||||
|
|
@ -270,6 +276,11 @@ async def _generate_ai_response(
|
|||
"""
|
||||
import time
|
||||
|
||||
logger.info(f"[AC-ASA-19] ========== RAG LAB FULL PROMPT ==========")
|
||||
logger.info(f"[AC-ASA-19] Prompt length: {len(prompt)}")
|
||||
logger.info(f"[AC-ASA-19] Prompt content:\n{prompt}")
|
||||
logger.info(f"[AC-ASA-19] ==========================================")
|
||||
|
||||
try:
|
||||
manager = get_llm_config_manager()
|
||||
client = manager.get_client()
|
||||
|
|
@ -300,33 +311,10 @@ async def _generate_ai_response(
|
|||
def _build_final_prompt(query: str, retrieval_results: list[dict]) -> str:
|
||||
"""
|
||||
Build the final prompt from query and retrieval results.
|
||||
Uses shared prompt configuration for consistency with orchestrator.
|
||||
"""
|
||||
if not retrieval_results:
|
||||
return f"""用户问题：{query}
|
||||
|
||||
未找到相关检索结果，请基于通用知识回答用户问题。"""
|
||||
|
||||
evidence_text = "\n".join([
|
||||
f"{i+1}. [Score: {hit['score']:.2f}] {hit['content'][:200]}{'...' if len(hit['content']) > 200 else ''}"
|
||||
for i, hit in enumerate(retrieval_results[:5])
|
||||
])
|
||||
|
||||
return f"""【系统指令】
|
||||
你是一名经验丰富的客服专员，名字叫“小N”。请遵循以下准则：
|
||||
1. 使用自然、口语化的中文，语气亲切、有礼貌，像在和人聊天一样。
|
||||
2. 不要使用列表、编号或过于机械的结构，而是用连贯的段落表达。
|
||||
3. 如果知识库信息不足，可以委婉地说明“我帮您查一下”、“我确认一下”，而不是直接说“我不知道”。
|
||||
4. 不要提及“作为AI”、“根据算法”等词，避免暴露机器身份。
|
||||
5. 适当使用语气词（如“呢”、“哦”、“哈”）、表情符号（根据品牌调性可选）和人性化的表达（如“我理解”、“您别着急”）。
|
||||
6. 回答长度适中，一般不超过3-5句话，如需详细说明可分段落，但保持易读。
|
||||
7. 如果问题涉及隐私或敏感信息，请引导客户联系人工客服（如“为了您的账户安全，建议您拨打官方电话...”）。
|
||||
|
||||
【知识库内容】
|
||||
{evidence_text}
|
||||
|
||||
【用户问题】
|
||||
{query}
|
||||
"""
|
||||
evidence_text = format_evidence_for_prompt(retrieval_results, max_results=5, max_content_length=500)
|
||||
return build_user_prompt_with_evidence(query, evidence_text)
|
||||
|
||||
|
||||
def _get_fallback_results(query: str) -> list[dict]:
|
||||
|
|
|
|||
|
|
@ -0,0 +1,114 @@
|
|||
"""
|
||||
Shared prompt templates for AI Service.
|
||||
Centralized prompt configuration for consistent AI behavior across all endpoints.
|
||||
"""
|
||||
|
||||
SYSTEM_PROMPT = """你是一名经验丰富的客服专员，名字叫"小N"。请遵循以下准则：
|
||||
像朋友一样聊天：使用自然、口语化的中文，语气亲切有礼貌，就像在和一位老朋友聊天。多使用"咱们"、"您看"、"是不是"这样拉近距离的表达。
|
||||
避免生硬罗列：介绍多个信息时，不要用"第一、第二"或分点列举，而是用流畅的段落把它们串联起来。比如："我们这边有几类课程挺受欢迎的，一个是……另外还有……您对哪方面更感兴趣呀？"
|
||||
主动引导对话：不要一次性把知识库里的所有内容都倒出来。可以先简单概括，然后问问客户具体想了解什么，比如："您是想提升职场英语，还是有留学方面的打算呢？"
|
||||
委婉处理未知：如果知识库里没有答案，别直接说"不知道"，可以说："您稍等，我帮您确认一下"、"这个问题我需要问问同事，马上给您回复"。
|
||||
隐藏机器身份：绝不要提及"作为AI"、"根据算法"等词，也避免使用过于官方的术语。就像普通人一样用"我"、"我们"。
|
||||
多用语气词和情感表达：适当加入"呢"、"哦"、"哈"、"呀"等语气词，根据情况使用表情符号（😊）。表达理解和关心，比如"我明白您的意思"、"别着急，我们一起看看"。
|
||||
保持简洁但有温度：回答一般控制在3-5句话，如果需要详细说明，可以分成小段落，但每段都要口语化，读起来不累。
|
||||
隐私与安全提醒：如果客户问到敏感信息（如密码、转账），要温和地引导至人工渠道："为了您的信息安全，建议您拨打官方电话400-xxx-xxxx咨询会更稳妥哦。"""
|
||||
|
||||
|
||||
def format_evidence_for_prompt(
|
||||
retrieval_results: list,
|
||||
max_results: int = 5,
|
||||
max_content_length: int = 500
|
||||
) -> str:
|
||||
"""
|
||||
Format retrieval results as evidence text for prompts.
|
||||
|
||||
Args:
|
||||
retrieval_results: List of retrieval hits. Can be:
|
||||
- dict format: {'content', 'score', 'source', 'metadata'}
|
||||
- RetrievalHit object: with .text, .score, .source, .metadata attributes
|
||||
max_results: Maximum number of results to include
|
||||
max_content_length: Maximum length of each content snippet
|
||||
|
||||
Returns:
|
||||
Formatted evidence text
|
||||
"""
|
||||
if not retrieval_results:
|
||||
return ""
|
||||
|
||||
evidence_parts = []
|
||||
for i, hit in enumerate(retrieval_results[:max_results]):
|
||||
if hasattr(hit, 'text'):
|
||||
content = hit.text
|
||||
score = hit.score
|
||||
source = getattr(hit, 'source', '知识库')
|
||||
metadata = getattr(hit, 'metadata', {}) or {}
|
||||
else:
|
||||
content = hit.get('content', '')
|
||||
score = hit.get('score', 0)
|
||||
source = hit.get('source', '知识库')
|
||||
metadata = hit.get('metadata', {}) or {}
|
||||
|
||||
if len(content) > max_content_length:
|
||||
content = content[:max_content_length] + '...'
|
||||
|
||||
nested_meta = metadata.get('metadata', {})
|
||||
source_doc = nested_meta.get('source_doc', source) if nested_meta else source
|
||||
category = nested_meta.get('category', '') if nested_meta else ''
|
||||
department = nested_meta.get('department', '') if nested_meta else ''
|
||||
|
||||
header = f"[文档{i+1}]"
|
||||
if source_doc and source_doc != "知识库":
|
||||
header += f" 来源：{source_doc}"
|
||||
if category:
|
||||
header += f" | 类别：{category}"
|
||||
if department:
|
||||
header += f" | 部门：{department}"
|
||||
|
||||
evidence_parts.append(f"{header}\n相关度：{score:.2f}\n内容：{content}")
|
||||
|
||||
return "\n\n".join(evidence_parts)
|
||||
|
||||
|
||||
def build_system_prompt_with_evidence(evidence_text: str) -> str:
|
||||
"""
|
||||
Build system prompt with knowledge base evidence.
|
||||
|
||||
Args:
|
||||
evidence_text: Formatted evidence from retrieval results
|
||||
|
||||
Returns:
|
||||
Complete system prompt
|
||||
"""
|
||||
if not evidence_text:
|
||||
return SYSTEM_PROMPT
|
||||
|
||||
return f"""{SYSTEM_PROMPT}
|
||||
|
||||
知识库参考内容：
|
||||
{evidence_text}"""
|
||||
|
||||
|
||||
def build_user_prompt_with_evidence(query: str, evidence_text: str) -> str:
|
||||
"""
|
||||
Build user prompt with knowledge base evidence (for single-message format).
|
||||
|
||||
Args:
|
||||
query: User's question
|
||||
evidence_text: Formatted evidence from retrieval results
|
||||
|
||||
Returns:
|
||||
Complete user prompt
|
||||
"""
|
||||
if not evidence_text:
|
||||
return f"""用户问题：{query}
|
||||
|
||||
未找到相关检索结果，请基于通用知识回答用户问题。"""
|
||||
|
||||
return f"""【系统指令】
|
||||
{SYSTEM_PROMPT}
|
||||
|
||||
【知识库内容】
|
||||
{evidence_text}
|
||||
|
||||
【用户问题】
|
||||
{query}"""
|
||||
|
|
@ -133,6 +133,13 @@ class OpenAIClient(LLMClient):
|
|||
body = self._build_request_body(messages, effective_config, stream=False, **kwargs)
|
||||
|
||||
logger.info(f"[AC-AISVC-02] Generating response with model={effective_config.model}")
|
||||
logger.info(f"[AC-AISVC-02] ========== FULL PROMPT TO AI ==========")
|
||||
for i, msg in enumerate(messages):
|
||||
role = msg.get("role", "unknown")
|
||||
content = msg.get("content", "")
|
||||
logger.info(f"[AC-AISVC-02] [{i}] role={role}, content_length={len(content)}")
|
||||
logger.info(f"[AC-AISVC-02] [{i}] content:\n{content}")
|
||||
logger.info(f"[AC-AISVC-02] ======================================")
|
||||
|
||||
try:
|
||||
response = await client.post(
|
||||
|
|
@ -213,6 +220,13 @@ class OpenAIClient(LLMClient):
|
|||
body = self._build_request_body(messages, effective_config, stream=True, **kwargs)
|
||||
|
||||
logger.info(f"[AC-AISVC-06] Starting streaming generation with model={effective_config.model}")
|
||||
logger.info(f"[AC-AISVC-06] ========== FULL PROMPT TO AI (STREAMING) ==========")
|
||||
for i, msg in enumerate(messages):
|
||||
role = msg.get("role", "unknown")
|
||||
content = msg.get("content", "")
|
||||
logger.info(f"[AC-AISVC-06] [{i}] role={role}, content_length={len(content)}")
|
||||
logger.info(f"[AC-AISVC-06] [{i}] content:\n{content}")
|
||||
logger.info(f"[AC-AISVC-06] ======================================")
|
||||
|
||||
try:
|
||||
async with client.stream(
|
||||
|
|
|
|||
|
|
@ -25,6 +25,7 @@ from typing import Any, AsyncGenerator
|
|||
from sse_starlette.sse import ServerSentEvent
|
||||
|
||||
from app.core.config import get_settings
|
||||
from app.core.prompts import SYSTEM_PROMPT, format_evidence_for_prompt
|
||||
from app.core.sse import (
|
||||
create_error_event,
|
||||
create_final_event,
|
||||
|
|
@ -41,16 +42,6 @@ from app.services.retrieval.base import BaseRetriever, RetrievalContext, Retriev
|
|||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
OPTIMIZED_SYSTEM_PROMPT = """你是学校智能客服助手，基于提供的知识库内容回答用户问题。
|
||||
|
||||
回答要求：
|
||||
1. 严格基于提供的知识库内容回答，不要编造信息
|
||||
2. 如果知识库中没有相关信息，明确告知用户并建议转人工或稍后重试
|
||||
3. 保持专业、友好的语气，回答简洁明了，突出重点
|
||||
4. 如果引用知识库内容，请注明来源（如：根据[文档1]...）
|
||||
5. 对于时效性问题，请提醒用户注意文档的有效期"""
|
||||
|
||||
|
||||
@dataclass
|
||||
class OrchestratorConfig:
|
||||
"""
|
||||
|
|
@ -59,7 +50,7 @@ class OrchestratorConfig:
|
|||
"""
|
||||
max_history_tokens: int = 4000
|
||||
max_evidence_tokens: int = 2000
|
||||
system_prompt: str = OPTIMIZED_SYSTEM_PROMPT
|
||||
system_prompt: str = SYSTEM_PROMPT
|
||||
enable_rag: bool = True
|
||||
use_optimized_retriever: bool = True
|
||||
|
||||
|
|
@ -408,32 +399,23 @@ class OrchestratorService:
|
|||
f"system_len={len(system_content)}, history_count={len(ctx.merged_context.messages) if ctx.merged_context else 0}"
|
||||
)
|
||||
logger.debug(f"[AC-AISVC-02] System prompt preview: {system_content[:500]}...")
|
||||
|
||||
logger.info(f"[AC-AISVC-02] ========== ORCHESTRATOR FULL PROMPT ==========")
|
||||
for i, msg in enumerate(messages):
|
||||
role = msg.get("role", "unknown")
|
||||
content = msg.get("content", "")
|
||||
logger.info(f"[AC-AISVC-02] [{i}] role={role}, content_length={len(content)}")
|
||||
logger.info(f"[AC-AISVC-02] [{i}] content:\n{content}")
|
||||
logger.info(f"[AC-AISVC-02] ==============================================")
|
||||
|
||||
return messages
|
||||
|
||||
def _format_evidence(self, retrieval_result: RetrievalResult) -> str:
|
||||
"""
|
||||
[AC-AISVC-17] Format retrieval hits as evidence text.
|
||||
Optimized format with source attribution and metadata.
|
||||
Uses shared prompt configuration for consistency.
|
||||
"""
|
||||
evidence_parts = []
|
||||
for i, hit in enumerate(retrieval_result.hits[:5], 1):
|
||||
metadata = hit.metadata or {}
|
||||
source = metadata.get("metadata", {}).get("source_doc", "知识库")
|
||||
category = metadata.get("metadata", {}).get("category", "")
|
||||
department = metadata.get("metadata", {}).get("department", "")
|
||||
|
||||
header = f"[文档{i}]"
|
||||
if source and source != "知识库":
|
||||
header += f" 来源：{source}"
|
||||
if category:
|
||||
header += f" | 类别：{category}"
|
||||
if department:
|
||||
header += f" | 部门：{department}"
|
||||
|
||||
evidence_parts.append(f"{header}\n相关度：{hit.score:.2f}\n内容：{hit.text}")
|
||||
|
||||
return "\n\n".join(evidence_parts)
|
||||
return format_evidence_for_prompt(retrieval_result.hits, max_results=5, max_content_length=500)
|
||||
|
||||
def _fallback_response(self, ctx: GenerationContext) -> str:
|
||||
"""
|
||||
|
|
|
|||
Loading…
Reference in New Issue