diff --git a/app.py b/app.py index 0640246..99bc344 100644 --- a/app.py +++ b/app.py @@ -321,7 +321,7 @@ def display_product_card_from_item(product: ProductItem) -> None: if product.price is not None: st.caption(f"¥{product.price:.2f}") - label_style = "⭐" if product.match_label == "Highly Relevant" else "✦" + label_style = "⭐" if product.match_label == "Relevant" else "✦" st.caption(f"{label_style} {product.match_label}") @@ -331,7 +331,7 @@ def render_search_result_block(result: SearchResult) -> None: Shows: - A styled header with query + match counts + quality_summary (if any) - - A grid of product cards (Highly Relevant first, then Partially Relevant; max 6) + - A grid of product cards (Relevant first, then Partially Relevant; max 6) """ summary_line = f'  · {result.quality_summary}' if result.quality_summary else '' header_html = ( @@ -339,7 +339,7 @@ def render_search_result_block(result: SearchResult) -> None: f'margin:8px 0 4px 0;background:#fafafa;">' f'' f'🔍 {result.query}' - f' · Highly Relevant {result.perfect_count} 件' + f' · Relevant {result.perfect_count} 件' f' · Partially Relevant {result.partial_count} 件' f'{summary_line}' f'' @@ -347,7 +347,7 @@ def render_search_result_block(result: SearchResult) -> None: st.markdown(header_html, unsafe_allow_html=True) # Perfect matches first, fall back to partials if none - perfect = [p for p in result.products if p.match_label == "Highly Relevant"] + perfect = [p for p in result.products if p.match_label == "Relevant"] partial = [p for p in result.products if p.match_label == "Partially Relevant"] to_show = (perfect + partial)[:6] if perfect else partial[:6] @@ -361,14 +361,20 @@ def render_search_result_block(result: SearchResult) -> None: display_product_card_from_item(product) -def render_message_with_refs(content: str, session_id: str) -> None: +def render_message_with_refs( + content: str, + session_id: str, + fallback_refs: Optional[dict] = None, +) -> None: """ Render an assistant message that may contain [SEARCH_REF:xxx] tokens. Text segments are rendered as markdown. [SEARCH_REF:xxx] tokens are replaced with full product card blocks - loaded from the global registry. + loaded from the global registry, or from fallback_refs (e.g. refs stored + with the message so they survive reruns / different workers). """ + fallback_refs = fallback_refs or {} # re.split with a capture group alternates: [text, ref_id, text, ref_id, ...] parts = SEARCH_REF_PATTERN.split(content) @@ -381,7 +387,7 @@ def render_message_with_refs(content: str, session_id: str) -> None: else: # ref_id segment ref_id = segment.strip() - result = global_registry.get(session_id, ref_id) + result = global_registry.get(session_id, ref_id) or fallback_refs.get(ref_id) if result: render_search_result_block(result) else: @@ -450,7 +456,9 @@ def display_message(message: dict): # Render message: expand [SEARCH_REF:xxx] tokens into product card blocks session_id = st.session_state.get("session_id", "") - render_message_with_refs(content, session_id) + render_message_with_refs( + content, session_id, fallback_refs=message.get("search_refs") + ) st.markdown("", unsafe_allow_html=True) @@ -671,13 +679,14 @@ def main(): tool_calls = result.get("tool_calls", []) debug_steps = result.get("debug_steps", []) - # Add assistant message + # Add assistant message (store search_refs so refs resolve after rerun) st.session_state.messages.append( { "role": "assistant", "content": response, "tool_calls": tool_calls, "debug_steps": debug_steps, + "search_refs": result.get("search_refs", {}), } ) diff --git a/app/agents/shopping_agent.py b/app/agents/shopping_agent.py index 033e62c..674d1bc 100644 --- a/app/agents/shopping_agent.py +++ b/app/agents/shopping_agent.py @@ -10,6 +10,8 @@ Architecture: import json import logging +import re +from datetime import datetime from pathlib import Path from typing import Any, Optional, Sequence @@ -33,7 +35,30 @@ logger = logging.getLogger(__name__) # 1. Guides multi-query search planning with explicit evaluate-and-decide loop # 2. Forbids re-listing product details in the final response # 3. Mandates [SEARCH_REF:xxx] inline citation as the only product presentation mechanism -SYSTEM_PROMPT = """ 角色定义 +SYSTEM_PROMPT = f"""角色定义 +你是我们店铺的一名专业的电商导购，是一个善于倾听、主动引导、懂得搭配的“时尚顾问”，通过有温度的对话，给用户提供有价值的信息，包括需求引导、方案推荐、搜索结果推荐，最终促成满意的购物决策或转化行为。 +作为我们店铺的一名专业的销售，除了本店铺的商品的推荐，你可以给用户提供有帮助的信息，但是不要虚构商品、提供本商店搜索结果以外的商品。 + +一些原则： +1. 价值提供与信息收集的原则： + 1. 优先价值提供：适时的提供有价值的信息，如商品推荐、穿搭建议、趋势信息，在推荐方向上有需求缺口、需要明确的重要信息时，要适时的做“信息收集”，引导式的澄清需求、提高商品发现的效率，形成“提供-反馈”的良性循环。 + 2. 意图判断-缺口大（比如品类或者使用人群都不能确定）：从“品类”、“场景”、“效果”等宽泛的意图切入，给出方案推荐 + 1-2个关键问题让用户选择；示例： + 1. 你想穿出哪种感觉？职场干练 松弛自在 活力元气 温柔知性 + 2. 平时通勤场合多吗？还是更喜欢生活化穿搭？ + 3. 意图判断-缺口小：直接检索+方案呈现，根据情况，可以考虑该方向下重要的决策因素（思考哪些维度最可能影响推荐结果），进行提议和问题收集，让用户既得到相关信息、又得到下一步的方向引导、同时也有机会修正或者细化诉求。 + 4. 选项驱动式澄清：推荐几个清晰的方向，呈现方案或商品搜索结果，再做澄清 + 5. 单轮对话最好只提一个问题，最多两个，禁止多问题堆叠。 + 6. 站在用户立场思考：比如询问用户期待的效果或感觉、使用的场合、想解决的问题，而不是询问具体的款式、参数，你需要将用户表达的需求翻译为具体可检索的商品特征（版型、材质、设计元素、风格标签等），并据此筛选商品、组织推荐逻辑。 +2. 如何使用make_search_products_tool： + 1. 可以生成多个query进行搜索：在需要搜索商品的时候，可以将需求分解为 2-4 个搜索查询，每个 query 聚焦一个明确的商品子类或搜索角度。 + 2. 可以根据搜索结果调整搜索策略：每次调用 search_products 后，工具会返回搜索结果的相关性的判断、以及搜索结果的topN的title，你需要决策是否要调整搜索策略，比如结果质量太差，可能需要调整搜索词、或者加大试探的query数量（不要超过3-5个）。结果太差的原因有可能是你生成的query不合理、请根据你看到的商品名称的构成组织搜索关键词。 +3. 在最终回复中使用 [SEARCH_REF:xxx] 内联引用搜索结果： + 1. 搜索工具会返回一个结果引用标识[SEARCH_REF:xxx]，撰写最终答复的时候请直接引用 [SEARCH_REF:xxx] ，系统会自动在该位置渲染对应的商品卡片列表，无需复述搜索结果。 + 2. 因为系统会自动将[SEARCH_REF:xxx]渲染为搜索结果，所以[SEARCH_REF:xxx]必须独占一行，且只在需要渲染该query完整的搜索结果时才进行引用，同一个结果不要重复引用。 +4. 今天是{datetime.now().strftime("%Y-%m-%d")}，所有与当前时间（比如天气、最新或即将发生的事件）相关的问题，都要使用web_search工具）。 +""" + +SYSTEM_PROMPT___2 = """ 角色定义 你是我们店铺的一名专业的电商导购，是一个善于倾听、主动引导、懂得搭配的“时尚顾问”，通过有温度的对话，给用户提供有价值的信息，包括需求引导、方案推荐、搜索结果推荐，最终促成满意的购物决策或转化行为。 作为我们店铺的一名专业的销售，除了本店铺的商品的推荐，你可以给用户提供有帮助的信息，但是不要虚构商品、提供本商店搜索结果以外的商品。 @@ -48,6 +73,7 @@ SYSTEM_PROMPT = """ 角色定义 1. 可以生成多个query进行搜索：在需要搜索商品的时候，可以将需求分解为 2-4 个搜索查询，每个 query 聚焦一个明确的商品子类或搜索角度。 2. 可以根据搜索结果调整搜索策略：每次调用 search_products 后，工具会返回搜索结果的相关性的判断、以及搜索结果的topN的title，你需要决策是否要调整搜索策略，比如结果质量太差，可能需要调整搜索词、或者加大试探的query数量（不要超过3-5个）。 3. 使用 [SEARCH_REF:xxx] 内联引用搜索结果：搜索工具会返回一个结果引用标识[SEARCH_REF:xxx]，撰写最终答复的时候可以直接引用将 [SEARCH_REF:xxx] ，系统会自动在该位置渲染对应的商品卡片列表，无需复述搜索结果。 + 4. 因为系统会自动将[SEARCH_REF:xxx]渲染为搜索结果，所以只在需要渲染该query完整的搜索结果时才进行引用，同一个结果不要重复引用。 """ @@ -81,9 +107,40 @@ def _extract_message_text(msg) -> str: return str(content) if content else "" +# 部分 API（如 DeepSeek）在 content 中返回 think 标签块，需去掉后只保留正式回复 +_RE_THINK_TAGS = re.compile(r".*?<\/think>", re.DOTALL | re.IGNORECASE) + + +def _extract_formal_reply(msg) -> str: + """ + 只截取大模型回复中的「正式结果」，去掉 thinking/reasoning 内容。 + - 若 content 为 list（如 Responses API）：只取 type 为 output_text/text 的块，跳过 reasoning。 + - 若 content 为 str：去掉 think 标签及其内容。 + """ + content = getattr(msg, "content", "") + if isinstance(content, list): + parts = [] + for block in content: + if not isinstance(block, dict): + continue + block_type = (block.get("type") or "").lower() + if block_type in ("reasoning",): + continue + text = block.get("text") or block.get("content") or "" + if text: + parts.append(text) + return "".join(str(p) for p in parts).strip() + if isinstance(content, str): + return _RE_THINK_TAGS.sub("", content).strip() + return str(content).strip() if content else "" + + def _message_for_log(msg: BaseMessage) -> dict: """Serialize a message for structured logging (content truncated).""" - text = _extract_message_text(msg) + if getattr(msg, "additional_kwargs", None) and "reasoning" in (msg.additional_kwargs or {}): + text = _extract_formal_reply(msg) or _extract_message_text(msg) + else: + text = _extract_message_text(msg) if len(text) > _LOG_CONTENT_MAX: text = text[:_LOG_CONTENT_MAX] + f"... [truncated, total {len(text)} chars]" out: dict[str, Any] = { @@ -106,13 +163,17 @@ class ShoppingAgent: def __init__(self, session_id: Optional[str] = None): self.session_id = session_id or "default" - llm_kwargs = dict( + llm_kwargs: dict[str, Any] = dict( model=settings.openai_model, temperature=settings.openai_temperature, api_key=settings.openai_api_key, ) if settings.openai_api_base_url: llm_kwargs["base_url"] = settings.openai_api_base_url + if getattr(settings, "openai_use_reasoning", False): + llm_kwargs["use_responses_api"] = True + effort = getattr(settings, "openai_reasoning_effort", "medium") or "medium" + llm_kwargs["model_kwargs"] = {"reasoning": {"effort": effort, "summary": "none"}} self.llm = ChatOpenAI(**llm_kwargs) @@ -246,7 +307,7 @@ class ShoppingAgent: final_state = self.graph.get_state(config) final_msg = final_state.values["messages"][-1] - response_text = _extract_message_text(final_msg) + response_text = _extract_formal_reply(final_msg) or _extract_message_text(final_msg) # Collect new SearchResults added during this turn registry_after = global_registry.get_all(self.session_id) @@ -292,7 +353,8 @@ class ShoppingAgent: if getattr(msg, "type", None) in ("system", "tool"): continue role = "user" if msg.type == "human" else "assistant" - result.append({"role": role, "content": _extract_message_text(msg)}) + content = _extract_formal_reply(msg) or _extract_message_text(msg) if role == "assistant" else _extract_message_text(msg) + result.append({"role": role, "content": content}) return result except Exception as e: logger.error(f"get_conversation_history error: {e}") diff --git a/app/config.py b/app/config.py index 3ce61b4..9940c8c 100644 --- a/app/config.py +++ b/app/config.py @@ -33,6 +33,9 @@ class Settings(BaseSettings): openai_vision_model: str = "qwen3-omni-flash" openai_temperature: float = 0.7 openai_max_tokens: int = 1000 + # 对话调用大模型时是否开启 thinking（需兼容 Responses API / reasoning 的模型，如 o1/o3/o4-mini） + openai_use_reasoning: bool = False + openai_reasoning_effort: str = "medium" # low | medium | high # Base URL for OpenAI-compatible APIs (e.g. Qwen/DashScope) # Qwen 北京: https://dashscope.aliyuncs.com/compatible-mode/v1 openai_api_base_url: Optional[str] = None diff --git a/app/search_registry.py b/app/search_registry.py index 48ffc3d..6db29ff 100644 --- a/app/search_registry.py +++ b/app/search_registry.py @@ -27,7 +27,7 @@ class ProductItem: vendor: Optional[str] = None image_url: Optional[str] = None relevance_score: Optional[float] = None - # LLM-assigned label: "Highly Relevant" | "Partially Relevant" | "Not Relevant" + # LLM-assigned label: "Relevant" | "Partially Relevant" | "Irrelevant" match_label: str = "Partially Relevant" tags: list = field(default_factory=list) specifications: list = field(default_factory=list) @@ -40,7 +40,7 @@ class SearchResult: Identified by ref_id (e.g. 'sr_3f9a1b2c'). Stores the query, LLM quality assessment, and the curated product list - (only "Highly Relevant" and "Partially Relevant" items — "Not Relevant" are discarded). + (only "Relevant" and "Partially Relevant" items — "Irrelevant" are discarded). """ ref_id: str diff --git a/app/tools/search_tools.py b/app/tools/search_tools.py index eae9878..1db6eff 100644 --- a/app/tools/search_tools.py +++ b/app/tools/search_tools.py @@ -2,7 +2,7 @@ Search Tools for Product Discovery - search_products is created via make_search_products_tool(session_id, registry). -- After search API, an LLM labels each result as Highly Relevant / Partially Relevant / Not Relevant; we count and +- After search API, an LLM labels each result as Relevant / Partially Relevant / Irrelevant; we count and store the curated list in the registry, return [SEARCH_REF:ref_id] + quality counts + top10 titles. """ @@ -74,7 +74,7 @@ def _assess_search_quality(query: str, raw_products: list) -> tuple[list[str], s product_text = "\n".join(lines) prompt = f"""评估以下搜索结果与用户查询的匹配程度，完成两件事： -1. 为每条结果打一个等级：Highly Relevant / Partially Relevant / Not Relevant。 +1. 为每条结果打一个等级：Relevant / Partially Relevant / Irrelevant。 2. 写一段 quality_summary（1–2 句话）：简要说明搜索结果主要包含哪些商品、是否基本满足搜索意图、整体匹配度如何。 用户查询：{query} @@ -82,10 +82,10 @@ def _assess_search_quality(query: str, raw_products: list) -> tuple[list[str], s 搜索结果（共 {n} 条）： {product_text} -等级说明：Highly Relevant=完全符合查询意图；Partially Relevant=基本相关（如品类等主需求匹配但部分属性不完全符合）；Not Relevant=不相关。 +等级说明：Relevant=完全符合查询意图；Partially Relevant=基本相关（如品类等主需求匹配但部分属性不完全符合）；Irrelevant=不相关。 请严格按以下 JSON 输出，仅输出 JSON，无其他内容： -{{"labels": ["Highly Relevant", "Partially Relevant", "Not Relevant", ...], "quality_summary": "你的1-2句总结"}} +{{"labels": ["Relevant", "Partially Relevant", "Irrelevant", ...], "quality_summary": "你的1-2句总结"}} labels 数组长度必须等于 {n}。""" try: @@ -93,7 +93,7 @@ labels 数组长度必须等于 {n}。""" resp = client.chat.completions.create( model=settings.openai_model, messages=[{"role": "user", "content": prompt}], - max_tokens=700, + max_tokens=1200, temperature=0.1, ) raw = resp.choices[0].message.content.strip() @@ -104,7 +104,7 @@ labels 数组长度必须等于 {n}。""" raw = raw.strip() data = json.loads(raw) labels = data.get("labels", []) - valid = {"Highly Relevant", "Partially Relevant", "Not Relevant"} + valid = {"Relevant", "Partially Relevant", "Irrelevant"} labels = [l if l in valid else "Partially Relevant" for l in labels] while len(labels) < n: labels.append("Partially Relevant") @@ -133,14 +133,14 @@ def make_search_products_tool( @tool def search_products(query: str, limit: int = 20) -> str: - """搜索商品库并做质量评估：LLM 为每条结果打等级（Highly Relevant / Partially Relevant / Not Relevant），返回引用与 top10 标题。 + """搜索商品库并做质量评估：LLM 为每条结果打等级（Relevant / Partially Relevant / Irrelevant），返回引用与 top10 标题。 Args: query: 自然语言商品描述 limit: 最多返回条数（1-20） Returns: - 【搜索完成】+ 结果引用 [SEARCH_REF:ref_id] + 质量情况（评估条数、Highly/Partially Relevant 数）+ results list（top10 标题） + 【搜索完成】+ 结果引用 [SEARCH_REF:ref_id] + 质量情况（评估条数、Relevant/Partially Relevant 数）+ results list（top10 标题） """ try: logger.info(f"[{session_id}] search_products: query={query!r} limit={limit}") @@ -176,13 +176,13 @@ def make_search_products_tool( ) labels, quality_summary = _assess_search_quality(query, raw_results) - perfect_count = sum(1 for l in labels if l == "Highly Relevant") + perfect_count = sum(1 for l in labels if l == "Relevant") partial_count = sum(1 for l in labels if l == "Partially Relevant") irrelevant_count = len(labels) - perfect_count - partial_count products: list[ProductItem] = [] for raw, label in zip(raw_results, labels): - if label not in ("Highly Relevant", "Partially Relevant"): + if label not in ("Relevant", "Partially Relevant"): continue products.append( ProductItem( @@ -229,7 +229,7 @@ def make_search_products_tool( return ( f"【搜索完成】query='{query}'\n" f"结果引用：[SEARCH_REF:{ref_id}]\n" - f"搜索结果质量情况：评估总条数{assessed_n}条，Highly Relevant {perfect_count} 条，Partially Relevant {partial_count} 条。\n" + f"搜索结果质量情况：评估总条数{assessed_n}条，Relevant {perfect_count} 条，Partially Relevant {partial_count} 条。\n" f"results list:\n{results_list}" ) @@ -251,7 +251,7 @@ def web_search(query: str) -> str: 触发场景： - 需要**外部知识**：流行趋势、品牌、搭配文化、节日习俗等 - - 需要**实时/及时信息**：当季流行元素、某地未来的天气 + - 需要**实时/及时信息**：所有与天气相关的问题、当季流行元素、某地近期或者未来的事件、所有依赖当前时间相关的信息 - 需要**宏观参考**：不同场合/国家的穿着建议、选购攻略 Args: @@ -369,7 +369,7 @@ def analyze_image_style(image_path: str) -> str: ], } ], - max_tokens=500, + max_tokens=800, temperature=0.3, ) -- libgit2 0.21.2