完善模型测试与参考文件历史能力

2026-07-02 17:37:35 +08:00
parent 8743a02110
commit d3530ab0b7
16 changed files with 697 additions and 32 deletions
@@ -43,12 +43,23 @@ def _ensure_json_content(text: str) -> dict:
    return {"content": [{"type": "text", "text": stripped}]}


+def _format_file_context(file_summaries: list[dict]) -> str:
+    file_blocks: list[str] = []
+    for item in file_summaries:
+        file_name = item.get("file_name") or "未命名文件"
+        summary = item.get("summary") or "文件内容为空。"
+        file_blocks.append(f"文件：{file_name}\n内容：\n{summary}")
+    return "\n\n".join(file_blocks)
+
+
 def _build_prompt(paragraph: Paragraph, file_summaries: list[dict] | None = None) -> tuple[str, str]:
    system_prompt = (
        "你是一个企业文档撰写助手。"
        "请严格输出 JSON，不要输出 JSON 之外的说明。"
        '格式为：{"content":[{"type":"text","text":"..."},{"type":"table","title":"...","headers":["..."],"rows":[["..."]]}]}。'
    )
+    if getattr(paragraph, "enable_reasoning", False):
+        system_prompt += "你可以先进行充分思考，再给出最终答案，但最终只输出要求的结果内容。"

    user_parts = [f"段落标题：{paragraph.title}"]
    if paragraph.content:
@@ -56,10 +67,7 @@ def _build_prompt(paragraph: Paragraph, file_summaries: list[dict] | None = None
    if paragraph.need_prompt and paragraph.prompt_text:
        user_parts.append(f"附加要求：{paragraph.prompt_text}")
    if file_summaries:
-        file_blocks = []
-        for item in file_summaries:
-            file_blocks.append(f"文件名：{item['file_name']}\n文件摘要：{item['summary']}")
-        user_parts.append("参考文件内容：\n" + "\n\n".join(file_blocks))
+        user_parts.append("参考文件内容：\n" + _format_file_context(file_summaries))
    user_parts.append(f"输出格式：{paragraph.output_format}")
    return system_prompt, "\n\n".join(user_parts)

@@ -136,6 +144,65 @@ async def _call_openai_compatible(model: AiModel, system_prompt: str, user_promp
    return AiCallResult(content=_ensure_json_content(text), raw_text=text, used_model=model.name)


+async def _stream_openai_compatible(model: AiModel, system_prompt: str, user_prompt: str):
+    api_key = decrypt_text(model.api_key_encrypted)
+    if not api_key:
+        raise RuntimeError("模型 API Key 不可用")
+
+    payload = {
+        "model": model.name,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ],
+        "temperature": 0.3,
+        "stream": True,
+    }
+    headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
+    last_error: Exception | None = None
+    for attempt in range(settings.AI_MAX_RETRIES):
+        try:
+            async with httpx.AsyncClient(timeout=settings.AI_REQUEST_TIMEOUT, trust_env=False) as client:
+                async with client.stream("POST", _normalize_openai_endpoint(model.api_endpoint), headers=headers, json=payload) as response:
+                    if response.status_code in (429, 500, 502, 503, 504):
+                        body = await response.aread()
+                        raise httpx.HTTPStatusError(
+                            f"上游模型流式响应异常: {response.status_code} - {body.decode('utf-8', errors='ignore')[:500]}",
+                            request=response.request,
+                            response=response,
+                        )
+                    response.raise_for_status()
+                    async for line in response.aiter_lines():
+                        if not line or not line.startswith("data:"):
+                            continue
+                        payload_line = line[5:].strip()
+                        if payload_line == "[DONE]":
+                            break
+                        try:
+                            chunk = json.loads(payload_line)
+                        except json.JSONDecodeError:
+                            continue
+                        delta_payload = chunk.get("choices", [{}])[0].get("delta", {})
+                        delta = delta_payload.get("content", "")
+                        reasoning = delta_payload.get("reasoning_content", "")
+                        if isinstance(delta, list):
+                            delta = "".join(
+                                item.get("text", "") if isinstance(item, dict) else str(item)
+                                for item in delta
+                            )
+                        if delta:
+                            yield delta
+                        if reasoning:
+                            yield reasoning
+                    return
+        except (httpx.TimeoutException, httpx.HTTPStatusError, httpx.NetworkError) as error:
+            last_error = error
+            if attempt == settings.AI_MAX_RETRIES - 1:
+                break
+            await asyncio.sleep(2 ** attempt)
+    raise RuntimeError(f"模型流式调用失败：{last_error}")
+
+
 async def _call_anthropic(model: AiModel, system_prompt: str, user_prompt: str) -> AiCallResult:
    api_key = decrypt_text(model.api_key_encrypted)
    if not api_key:
@@ -162,8 +229,79 @@ async def _call_anthropic(model: AiModel, system_prompt: str, user_prompt: str)
    return AiCallResult(content=_ensure_json_content(text), raw_text=text, used_model=model.name)


+async def _stream_anthropic(model: AiModel, system_prompt: str, user_prompt: str):
+    api_key = decrypt_text(model.api_key_encrypted)
+    if not api_key:
+        raise RuntimeError("模型 API Key 不可用")
+
+    payload = {
+        "model": model.name,
+        "max_tokens": 2048,
+        "system": system_prompt,
+        "messages": [{"role": "user", "content": user_prompt}],
+        "stream": True,
+    }
+    headers = {
+        "x-api-key": api_key,
+        "anthropic-version": "2023-06-01",
+        "content-type": "application/json",
+    }
+    last_error: Exception | None = None
+    for attempt in range(settings.AI_MAX_RETRIES):
+        try:
+            async with httpx.AsyncClient(timeout=settings.AI_REQUEST_TIMEOUT, trust_env=False) as client:
+                async with client.stream("POST", _normalize_anthropic_endpoint(model.api_endpoint), headers=headers, json=payload) as response:
+                    if response.status_code in (429, 500, 502, 503, 504):
+                        body = await response.aread()
+                        raise httpx.HTTPStatusError(
+                            f"上游模型流式响应异常: {response.status_code} - {body.decode('utf-8', errors='ignore')[:500]}",
+                            request=response.request,
+                            response=response,
+                        )
+                    response.raise_for_status()
+                    async for line in response.aiter_lines():
+                        if not line or not line.startswith("data:"):
+                            continue
+                        payload_line = line[5:].strip()
+                        if payload_line == "[DONE]":
+                            break
+                        try:
+                            chunk = json.loads(payload_line)
+                        except json.JSONDecodeError:
+                            continue
+                        if chunk.get("type") == "content_block_delta":
+                            delta = chunk.get("delta", {}).get("text", "")
+                            if delta:
+                                yield delta
+                    return
+        except (httpx.TimeoutException, httpx.HTTPStatusError, httpx.NetworkError) as error:
+            last_error = error
+            if attempt == settings.AI_MAX_RETRIES - 1:
+                break
+            await asyncio.sleep(2 ** attempt)
+    raise RuntimeError(f"模型流式调用失败：{last_error}")
+
+
 async def call_ai(paragraph: Paragraph, model: AiModel, file_summaries: list[dict] | None = None) -> AiCallResult:
    system_prompt, user_prompt = _build_prompt(paragraph, file_summaries)
    if model.api_format == "anthropic":
        return await _call_anthropic(model, system_prompt, user_prompt)
    return await _call_openai_compatible(model, system_prompt, user_prompt)
+
+
+def build_test_stream_prompt(paragraph: Paragraph, file_summaries: list[dict] | None = None) -> tuple[str, str]:
+    system_prompt = "你是一个企业文档撰写助手。请直接输出适合预览的正文内容或 Markdown 表格，不要输出 JSON。"
+    if getattr(paragraph, "enable_reasoning", False):
+        system_prompt += "你可以先进行充分思考，再持续输出最终可展示的内容。"
+    _, user_prompt = _build_prompt(paragraph, file_summaries or [])
+    return system_prompt, user_prompt
+
+
+async def stream_ai_preview(paragraph: Paragraph, model: AiModel, file_summaries: list[dict] | None = None):
+    system_prompt, user_prompt = build_test_stream_prompt(paragraph, file_summaries)
+    if model.api_format == "anthropic":
+        async for chunk in _stream_anthropic(model, system_prompt, user_prompt):
+            yield chunk
+        return
+    async for chunk in _stream_openai_compatible(model, system_prompt, user_prompt):
+        yield chunk
@@ -63,10 +63,12 @@ def summarize_file_bytes(file_name: str, content: bytes) -> str:
        return _decode_text(content)[:4000]
    if suffix == ".csv":
        return _summarize_csv(content)
-    if suffix in {".xlsx", ".xls"}:
+    if suffix in {".xlsx", ".xls", ".xlsm"}:
        return _summarize_excel(content, suffix)
    if suffix == ".docx":
        return _summarize_docx(content)
+    if suffix == ".doc":
+        return "当前暂不支持直接解析 .doc 旧版 Word 文件正文，建议先另存为 .docx 后再上传。"
    if suffix == ".pdf":
        return _summarize_pdf(content)
    return f"暂不支持解析该文件内容：{file_name}"
@@ -117,6 +117,7 @@ async def run_generation(document_id: int, template_id: int):
                        if model is None:
                            content = build_mock_content(paragraph)
                        else:
+                            setattr(paragraph, "enable_reasoning", bool(model.enable_reasoning))
                            result_data = await call_ai(paragraph, model)
                            content = result_data.content
                        status = "success"