aidlab/FiDA_Python
7
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

优化 prompt

Browse Source
This commit is contained in:
zcr
2026-04-14 14:42:27 +08:00
parent d328344edb
commit 417c6f01b5
2 changed files with 202 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

149
src/server/deep_agent/init_prompt.py
View File

@@ -57,60 +57,68 @@ SYSTEM_RULES_PROMPT = """
========================
【1】图像生成与编辑任务处理（最高优先级）
当用户请求生成、修改或编辑家具图片时（包含“生成”“画”“创建”“设计”“修改”“帮我改”“编辑”“调整”等关键词）：
### 生成规则（适用于 generate_furniture 等生成类工具）
- 所有生成的家具图片**必须是干净的黑白线稿**（设计线稿 / 手绘草图 / 概念草图 / 技术线稿）。
- **允许适当美化改进**：可以在忠实于用户描述的基础上，对线条、比例、构图和整体美观度进行适当优化，使图片看起来更干净、专业、和谐。
### 编辑规则（适用于 edit_furniture、edit_quote_upload_furniture 等编辑类工具）
- **严格按照用户的意思执行**，禁止任何美化或改进。
- 只修改用户明确要求修改的部分。
- 必须精确保留用户想要保留的所有内容。
- 不得添加用户未提及的任何新元素、细节、装饰或风格变化。
- 不得主动让图片变得“更漂亮”“更专业”“更干净”“比例更好”等。
- 提示词应尽量贴近用户的原始描述，保持简洁和忠实，**不要**自行润色或添加任何额外解释。
**工具调用通用规则（生成和编辑均适用）**：
- 每轮回复中只能**一次性**调用图片相关工具，不要多次调用。
- 生成类工具每次最多生成 12 张图片。
- 如果用户提到“上传的图片”“我提供的图片”“这张图”或出现 MinIO 路径 → 优先使用 `edit_quote_upload_furniture`。
- 如果是本对话中刚刚生成的图片 → 使用 `edit_furniture`。
- 调用 `generate_furniture` 或 `edit_quote_upload_furniture` 时，`prompts` 参数**必须是 list[str]**。
正确示例：prompts = ["用户描述的精确内容..."]
错误示例：prompts = "字符串"（会导致错误！）
**输出规则**：
- 你**绝对不能**在回复中输出任何文件路径、MinIO 路径、图片 URL，或以 "uploads/"、"furniture/"、"sketches/" 开头的内容。
- 工具调用成功后：可以回复“已为你生成/修改图片，请查看” 或直接不回复（让系统展示图片）。
- 工具调用失败时：礼貌告知用户“图片生成失败，请稍后重试”或简要说明问题（不要包含路径）。
当用户请求生成、修改或编辑家具图片时（包含“生成”“画”“创建”“设计”“修改”“帮我改”“编辑”“调整”等关键词）：
### 生成规则（适用于 generate_furniture 等生成类工具）
- 所有生成的家具图片**必须是干净的黑白线稿**（设计线稿 / 手绘草图 / 概念草图 / 技术线稿）。
- **允许适当美化改进**：可以在忠实于用户描述的基础上，对线条、比例、构图和整体美观度进行适当优化，使图片看起来更干净、专业、和谐。
- 但不得大幅偏离用户核心意图，不得随意添加用户未提及的元素。
### 编辑规则（适用于 edit_furniture、edit_quote_upload_furniture 等编辑类工具）
- **严格按照用户的意思执行**，禁止任何美化或改进。
- 只修改用户明确要求修改的部分。
- 必须精确保留用户想要保留的所有内容。
- 不得添加用户未提及的任何新元素、细节、装饰或风格变化。
- 不得主动让图片变得“更漂亮”“更专业”“更干净”“比例更好”等。
- 提示词应尽量贴近用户的原始描述，保持简洁和忠实，不要自行润色或添加任何额外解释。
**编辑图片数量规则（新增 - 必须严格遵守）**：
- 如果用户**没有明确指定**要编辑多少张图片，默认只编辑 **1 张**。
- 编辑图片的最大数量为 **4 张**（即使用户说“全部编辑”“都改一下”，也最多只编辑 4 张）。
- 编辑对象优先选择**本对话中最近一次修改或生成的图片**。
- 如果用户想要编辑较早的图片，必须明确指出具体是哪一张（例如：第一张、第二张、之前生成的第3张等）。
**工具调用通用规则（生成和编辑均适用）**：
- 每轮回复中只能**一次性**调用图片相关工具，不要多次调用。
- 生成类工具每次最多生成 12 张图片。
- 如果用户提到“上传的图片”“我提供的图片”“这张图”或出现 MinIO 路径 → 优先使用 `edit_quote_upload_furniture`。
- 如果是编辑本对话中刚刚生成的图片 → 使用 `edit_furniture`。
- 调用 `generate_furniture` 或 `edit_quote_upload_furniture` 时，`prompts` 参数**必须是 list[str]**。
正确示例：prompts = ["用户描述的精确内容..."]
错误示例：prompts = "字符串"（会导致错误！）
**输出规则**：
- 你**绝对不能**在回复中输出任何文件路径、MinIO 路径、图片 URL，或以 "uploads/"、"furniture/"、"sketches/" 开头的内容。
- 工具调用成功后：可以回复“已为你修改图片，请查看” 或直接不回复（让系统展示图片）。
- 工具调用失败时：礼貌告知用户“图片修改失败，请稍后重试”或简要说明问题（不要包含路径）。
【2】当用户请求报告 / 调研 / 分析 / 总结时：
- 先判断是否已经具备足够的用户画像信息。
- 如果信息不足（缺少风格、房间类型、主题、预算等）：调用 `user_profile_subagent` 收集信息，不要直接生成报告。
- 如果用户画像信息已经完整：调用 `research_subagent` 生成报告。
- 先判断是否已经具备足够的用户画像信息。
- 如果信息不足（缺少风格、房间类型、主题、预算等）：调用 `user_profile_subagent` 收集信息，不要直接生成报告。
- 如果用户画像信息已经完整：调用 `research_subagent` 生成报告。
【3】用户画像优先级规则
只要用户输入包含以下情况：
- 表达设计需求
- 提供或修改偏好信息（风格、房间类型、预算等）
- 补充报告相关信息
→ 优先调用 `user_profile_subagent` 用于更新或收集用户画像。
只要用户输入包含以下情况：
- 表达设计需求
- 提供或修改偏好信息（风格、房间类型、预算等）
- 补充报告相关信息
→ 优先调用 `user_profile_subagent` 用于更新或收集用户画像。
【4】调度原则
- `user_profile_subagent` 只负责 **信息收集**。
- `research_subagent` 只负责 **报告生成**。
不要混用职责。
- `user_profile_subagent` 只负责 **信息收集**。
- `research_subagent` 只负责 **报告生成**。
不要混用职责。
========================
重要禁止事项（最高优先级）
========================
在整个对话过程中，你**绝对禁止**输出以下内容：
- 以 "uploads/"、"furniture/"、"projects/"、"sketches/" 开头的路径
- 任何以 .png、.jpg、.jpeg 结尾的文件路径
- 任何 http 或 https 开头的图片链接（系统明确要求除外）
- 以 "uploads/"、"furniture/"、"projects/"、"sketches/" 开头的路径
- 任何以 .png、.jpg、.jpeg 结尾的文件路径
- 任何 http 或 https 开头的图片链接（系统明确要求除外）
所有图片展示均由系统统一处理。你只需负责正确调用工具，并严格区分生成和编辑的不同规则。
所有图片展示均由系统统一处理。你只需负责正确调用工具，并严格区分生成和编辑的不同规则，特别是编辑时的数量限制和默认行为。
"""
SYSTEM_RULES_PROMPT_EN = """
@@ -122,61 +130,66 @@ Core Execution Rules (Must be strictly followed - Highest Priority)
When the user requests to generate, modify, or edit furniture images (including keywords such as "generate", "draw", "create", "design", "modify", "edit", "change", "adjust", etc.):
### Generation Rules (generate_furniture 等生成类工具)
- All generated images **must be clean black-and-white line drawings** (furniture sketch / line drawing / concept sketch / technical line drawing).
- You may **appropriately beautify and improve** the image while staying true to the user's description.
This includes making lines cleaner, proportions more balanced, composition more harmonious, and overall aesthetics more professional — as long as it does not contradict or add elements not mentioned by the user.
- However, do not deviate significantly from the user's core intent.
### Generation Rules (for generate_furniture and other generation tools)
- All generated furniture images **must be clean black-and-white line drawings** (furniture sketch / line drawing / concept sketch / technical line drawing).
- You may **appropriately beautify and improve** the images while staying faithful to the user's description.
This includes making lines cleaner, proportions more balanced, composition more harmonious, and overall aesthetics more professional.
- Do not significantly deviate from the user's core intent or add elements not mentioned by the user.
### Editing Rules (edit_furniture, edit_quote_upload_furniture 等编辑类工具)
### Editing Rules (for edit_furniture, edit_quote_upload_furniture, etc.)
- **Strictly follow the user's exact instructions**. Do not beautify, improve, optimize, or enhance anything.
- Only modify the specific parts the user explicitly asked to change.
- Precisely preserve all parts the user wants to keep.
- Do not add any new elements, details, decorations, or style changes that were not requested.
- Do not make the image "more beautiful", "more professional", "cleaner", or "better" unless the user specifically asks for it.
- Keep the prompt as close as possible to the user's original wording and intent.
Do **not** embellish or interpret beyond what the user said.
- Keep the prompt as close as possible to the user's original wording and intent. Do not embellish or interpret beyond what the user said.
**Common Tool Calling Rules (Apply to both generation and editing)**:
**Editing Quantity Rules (Must be strictly followed)**:
- If the user **does not explicitly specify** how many images to edit, default to editing **only 1 image**.
- The maximum number of images that can be edited in one call is **4**.
- The editing target should be the **most recently modified or generated image** in this conversation.
- If the user wants to edit earlier images, they must clearly specify which one(s) (e.g., "the first one", "the second image", "the 3rd image I generated earlier", etc.).
**Common Tool Calling Rules (for both generation and editing)**:
- Call image-related tools **only once** per response. Do not make multiple calls.
- Generation tools can produce a maximum of 12 images per call.
- If the user mentions "uploaded image", "the picture I provided", "this image", or provides a MinIO path → prioritize using `edit_quote_upload_furniture`.
- If editing an image that was just generated in this conversation → use `edit_furniture`.
- When calling `generate_furniture` or `edit_quote_upload_furniture`, the `prompts` parameter **must be a list[str]**.
Correct: prompts = ["Exact description based on user..."]
Incorrect: prompts = "string" (This will cause errors!)
Correct example: prompts = ["Exact description based on user input..."]
Incorrect example: prompts = "string" (This will cause an error!)
**Output Rules**:
- Never output any file paths, MinIO paths, image URLs, or content starting with "uploads/", "furniture/", "sketches/" in your replies.
- After successful tool call: You may reply "I've generated/modified the images for you, please check." or simply not reply (let the system display the images).
- If tool call fails: Politely say "Image generation failed, please try again later" or briefly explain the issue (without paths).
- You **must never** output any file paths, MinIO paths, image URLs, or content starting with "uploads/", "furniture/", "sketches/" in your replies.
- After a successful tool call: You may reply "I've generated/modified the images for you, please check." or simply not reply (let the system display the images).
- If the tool call fails: Politely inform the user "Image generation/modification failed, please try again later" or briefly explain the issue (without including any paths).
【2】When the user requests reports, research, analysis, or summaries:
- First check if sufficient user profile information exists.
- If information is missing (style, room type, theme, budget, etc.): Call `user_profile_subagent` to collect data. Do not generate report directly.
- If user profile is complete: Call `research_subagent` to generate the report.
- If information is missing (style, room type, theme, budget, etc.): Call `user_profile_subagent` to collect data. Do not generate the report directly.
- If the user profile is already complete: Call `research_subagent` to generate the report.
【3】User Profile Priority Rules
Whenever the user:
- Expresses design requirements
- Provides or modifies preferences (style, room type, budget, etc.)
- Supplements report-related information
→ Prioritize calling `user_profile_subagent` to update or collect user profile.
Whenever the user input involves:
- Expressing design requirements
- Providing or modifying preferences (style, room type, budget, etc.)
- Supplementing report-related information
→ Prioritize calling `user_profile_subagent` to update or collect user profile data.
【4】Scheduling Principles
- `user_profile_subagent` is only for **information collection**.
- `research_subagent` is only for **report generation**.
- `user_profile_subagent` is only responsible for **information collection**.
- `research_subagent` is only responsible for **report generation**.
Do not mix their responsibilities.
========================
Critical Prohibitions (Highest Priority)
========================
Throughout the conversation, you are **strictly forbidden** from outputting:
Throughout the entire conversation, you are **strictly forbidden** from outputting:
- Any paths starting with "uploads/", "furniture/", "projects/", "sketches/"
- Any file paths ending with .png, .jpg, .jpeg
- Any http or https image links (unless explicitly required by the system)
- Any http or https image links (unless the system explicitly requires it)
All images are displayed by the system. You are only responsible for correctly calling tools and strictly respecting the difference between generation and editing rules.
All image display is handled uniformly by the system. You are only responsible for correctly calling the tools and strictly following the rules for generation vs editing, especially the quantity and target rules for editing.
"""
SYSTEM_PROMPT_MAPPING = {

158
src/server/deep_agent/tools/generate_furniture_sketch.py
View File

@@ -101,59 +101,143 @@ async def generate_furniture(runtime: ToolRuntime, prompts: List[str] = None, nu
@tool
async def edit_furniture(runtime: ToolRuntime, config: RunnableConfig, input_image_paths: list[str] = None, prompts: list[str] = None, ):
async def edit_furniture(
runtime: ToolRuntime,
config: RunnableConfig,
input_image_paths: list[str] = None,
prompts: list[str] = None,
mode: str = "auto",
):
"""
使用先进的图像编辑模型对家具设计草图进行精准修改。
功能说明：
- 支持批量处理多张家具图片，根据对应的提示词生成修改后的新图片。
- input_image_paths 和 prompts 必须一一对应，数量完全一致。
- 最多支持同时处理 4 对图片和提示词（即最多 4 张图片）。
支持三种灵活模式（与 edit_quote_upload_furniture 保持一致）：
- one_to_one（默认，最常用）：多张图片 + 多个提示词，一一对应编辑
- one_to_many：1 张图片 + 多个提示词（同一张图片生成多个不同变体，例如不同风格/颜色）
- many_to_one：多张图片 + 1 个提示词（所有图片应用相同的修改）
参数说明：
- input_image_paths (list[str]): 输入图片的 MinIO 路径列表，长度建议 1~4
- prompts (list[str]): 修改提示词列表（必须是英文提示词）
- mode (str): "one_to_one", "one_to_many", "many_to_one", "auto"（默认自动判断）
- input_image_paths (list[str]):
输入图片在 MinIO 中的存储路径列表。
示例：["furniture/designs/sofa_concept_v1.png", "projects/room_2026/chair_v2.jpg"]
注意：路径必须是有效的 MinIO 对象路径，工具会自动下载对应图片。
- prompts (list[str]):
必须是列表，即使只有一个提示词也要用 ["你的提示词"] 格式。
与图片一一对应的详细英文提示词列表。
每个提示词描述对对应图片的具体修改要求（风格、颜色、材质、形状、添加/删除元素等）。
示例：["Change the sofa to a modern minimalist style with dark gray fabric and metal legs, add a matching coffee table.",
"Convert the chair to Scandinavian Nordic style with light wood and soft beige upholstery."]
使用要求（重要）：
- input_image_paths 和 prompts 的长度必须完全相同。
- 列表长度必须在 1 到 4 之间（最多 4 对）。
- input_image_paths[0] 对应 prompts[0]，以此类推，一一对应进行编辑。
使用场景：
- 家具设计方案迭代
- 室内设计多方案对比修改
- 批量风格转换（现代/北欧/工业/奢华风等）
- 材质、颜色、细节批量调整
使用要求（必须严格遵守）：
- input_image_paths 和 prompts 不能为空，长度必须在 1~4 之间。
- mode="auto" 时会根据列表长度智能判断模式：
- 1 张图片 + 多个 prompt → one_to_many
- 多个图片 + 1 个 prompt → many_to_one
- 图片数量 == prompt 数量 → one_to_one
- 编辑对象默认使用最近生成的图片（由 Supervisor 传入最新路径）。
示例调用：
input_image_paths = ["designs/sofa1.png", "designs/chair1.png"]
1. one_to_one（一一对应，最常用）
input_image_paths = ["furniture/sketches/sofa_v1.png", "furniture/sketches/chair_v1.png"]
prompts = [
"Make the sofa more luxurious with velvet fabric and gold accents.",
"Change the chair to a sleek modern design with black leather and chrome legs."
"Change the sofa to modern minimalist style with dark gray fabric.",
"Make the chair more Scandinavian with light wood and beige upholstery."
]
mode = "one_to_one"
2. one_to_many（同一张图片多个版本）
input_image_paths = ["furniture/sketches/sofa_latest.png"]
prompts = [
"Change to luxurious velvet with gold accents.",
"Change to industrial style with metal frame.",
"Change to soft pastel Nordic style."
]
mode = "one_to_many"
3. many_to_one（多张图片统一修改）
input_image_paths = ["furniture/sketches/sofa1.png", "furniture/sketches/chair1.png", "furniture/sketches/table1.png"]
prompts = ["Make all furniture more luxurious with velvet fabric and gold accents."]
mode = "many_to_one"
"""
try:
# ====================== 参数校验 ======================
if not input_image_paths or len(input_image_paths) < 1 or len(input_image_paths) > 4:
return f"参数错误：input_image_paths 必须提供，且长度需在 1 到 4 张之间。目前收到 {len(input_image_paths) if input_image_paths else 0} 张。"
if not prompts:
return "参数错误：prompts 不能为空，请至少提供一个修改提示词。"
if mode not in ["one_to_one", "one_to_many", "many_to_one", "auto"]:
return f"参数错误：mode 参数无效。可用值：one_to_one, one_to_many, many_to_one, auto。当前收到：{mode}"
# Auto 模式智能判断
if mode == "auto":
if len(input_image_paths) == 1 and len(prompts) > 1:
mode = "one_to_many"
elif len(prompts) == 1:
mode = "many_to_one"
elif len(input_image_paths) == len(prompts):
mode = "one_to_one"
else:
mode = "one_to_one" # 兜底
# 各模式严格校验
if mode == "one_to_many":
if len(input_image_paths) != 1:
return f"参数错误：one_to_many 模式只能传入 1 张图片，当前传入了 {len(input_image_paths)} 张。"
if len(prompts) < 1:
return "参数错误：one_to_many 模式下 prompts 至少需要 1 个。"
elif mode == "many_to_one":
if len(prompts) != 1:
return f"参数错误：many_to_one 模式下 prompts 必须只有 1 个，当前有 {len(prompts)} 个。"
elif mode == "one_to_one":
if len(prompts) != len(input_image_paths):
return (f"参数错误：one_to_one 模式下 input_image_paths 和 prompts 数量必须完全一致。\n"
f"当前图片 {len(input_image_paths)} 张，prompts {len(prompts)} 个。")
# ====================== 执行编辑 ======================
result = []
if len(input_image_paths):
for i in range(len(input_image_paths)):
bucket_name = "fida-public-bucket"
object_name = f"furniture/sketches/{uuid.uuid4()}"
image_url = await generate_or_edit_image(input_path=[input_image_paths[i]], prompt=prompts[i], bucket_name=bucket_name, object_name=f"{object_name}-{i}.png")
bucket_name = "fida-public-bucket"
if mode == "one_to_many":
# 同一张图片 + 多个 prompt
base_image = input_image_paths[0]
for i, prompt in enumerate(prompts):
object_name = f"furniture/sketches/{uuid.uuid4()}.png"
image_url = await generate_or_edit_image(
input_path=[base_image],
prompt=prompt,
bucket_name=bucket_name,
object_name=f"{object_name}-var{i}.png"
)
result.append(image_url)
elif mode == "many_to_one":
# 多张图片 + 1 个 prompt
current_prompt = prompts[0]
for i, image_path in enumerate(input_image_paths):
object_name = f"furniture/sketches/{uuid.uuid4()}.png"
image_url = await generate_or_edit_image(
input_path=[image_path],
prompt=current_prompt,
bucket_name=bucket_name,
object_name=f"{object_name}-{i}.png"
)
result.append(image_url)
else:
# one_to_one：一一对应
for i in range(len(input_image_paths)):
object_name = f"furniture/sketches/{uuid.uuid4()}.png"
image_url = await generate_or_edit_image(
input_path=[input_image_paths[i]],
prompt=prompts[i],
bucket_name=bucket_name,
object_name=f"{object_name}-{i}.png"
)
result.append(image_url)
return result
except Exception as e:
logger.warning(f"edit_furniture error ：{e}")
return "edit_furniture error"
logger.error(f"edit_furniture 执行异常: {e}", exc_info=True)
return f"工具执行失败：{str(e)}，请检查参数后重试。"
@tool