diff --git a/README.md b/README.md index 411ac99..b6ab99e 100644 --- a/README.md +++ b/README.md @@ -10,7 +10,9 @@ - 同比、环比、自定义等长月份范围；缺基期禁止比较，缺成本显示不可用，缺月份趋势断开。 - 总览、异常追踪、贡献分解、自定义下钻层级、面包屑、原始行证据分页、CSV / XLSX 证据导出。 - 个人设置保存/取消/预览与乐观锁；调查备注与状态持久化；账号启停、角色及科室范围、操作审计。 -- 本地规则助手；可选 OpenAI Responses 结构化意图接口，用户同意后仅发送问题与维度/当前筛选，查询仍受后端权限约束。 +- 财务助手：规则摘要模式，或使用页面配置的模型基于权限内聚合结果生成动态回答并支持连续追问；回答标注“模型回答 · 实际模型”或“规则摘要”，普通问答不改变分析视图。 +- 对话新增分析列：助手追问口径、展示列定义预览，用户确认后保存到定义库并加入当前分析表；确认前不写入，同名替换保留原位置，最多 10 列。 +- 财务助手与公式工作簿的列设计共用组织的页面模型配置（Base URL、Key、模型），保留禁用时不回退、无数据库配置时才使用环境变量的语义。 当前原附件没有医生、成本和连续月度数据，不会生成虚构的医生排名、ROI 或环比结论。毛利是根据产品分类毛利率计算的贡献值，不等同于已审计净利润。 @@ -119,7 +121,7 @@ REFERENCE_DATASET_NAME='分析数据_原稿(1).xlsx' npm run test:e2e ## AI 与跨平台 -默认 AI_ENABLED=0。启用外部模型有两种方式，数据库配置优先：全权限管理员在页面「管理 → AI 模型」配置服务商（OpenAI 或本地 OpenAI 兼容服务如 Ollama / LM Studio / vLLM）、Base URL、API Key 与模型，可从 `/v1/models` 一键获取本地服务模型列表多选一；或配置环境变量 `AI_ENABLED=1`、`OPENAI_API_KEY`、`OPENAI_MODEL`（可选 `OPENAI_BASE_URL`）。均需完成机构数据处理审批。密钥只在后端：读取接口不回传密钥原文，配置修改记入审计（不含密钥）。用户自由输入的问题也可能含敏感信息，不能仅凭不发送原始行就认定无隐私风险。模型不能提交 SQL、修改金额或修改账号；输出只允许受控意图，异常/拒答关闭执行。 +默认 AI_ENABLED=0。启用外部模型有两种方式，数据库配置优先：全权限管理员在页面「管理 → AI 模型」配置服务商（OpenAI 或本地 OpenAI 兼容服务如 Ollama / LM Studio / vLLM）、Base URL、API Key 与模型，可从 `/v1/models` 一键获取本地服务模型列表多选一；或配置环境变量 `AI_ENABLED=1`、`OPENAI_API_KEY`、`OPENAI_MODEL`（可选 `OPENAI_BASE_URL`）。财务助手与公式工作簿的列设计读取同一份组织生效配置：数据库配置显式禁用时不回退环境变量，只有未保存数据库配置时才使用环境变量。用户勾选同意后，本轮对话、当前筛选、指标定义与权限范围内的聚合结果会发送到所配模型服务；不发送原始数据行、客户编码、源文件或密钥。回答标注“模型回答 · 实际模型名”或“规则摘要”，空输出、截断或请求失败返回明确错误，不伪装成功。对话新增列仅允许全范围 admin / analyst：模型只生成候选定义，须经服务端表达式与可用指标校验、并由用户确认后才保存；不接受 SQL、代码或任意公式文本。均需完成机构数据处理审批。密钥只在后端：读取接口不回传密钥原文，配置修改记入审计（不含密钥）。用户自由输入的问题也可能含敏感信息，不能仅凭不发送原始行就认定无隐私风险。模型不能提交 SQL、修改金额或修改账号；输出只允许受控意图，异常/拒答关闭执行。 前端为标准 Web，API 无平台耦合，可由后续 PWA、Tauri 或 Capacitor 客户端复用。当前没有打包桌面/移动客户端，也不在浏览器离线缓存财务数据。 diff --git a/finance/ai.py b/finance/ai.py index eaee881..931b04e 100644 --- a/finance/ai.py +++ b/finance/ai.py @@ -4,6 +4,8 @@ API Key 只存后端，读取接口只做脱敏回显，绝不把密钥下发到浏览器。 """ +import time + import httpx from django.conf import settings from django.http import JsonResponse @@ -13,6 +15,10 @@ from .views import APIError, audit, body, endpoint PROVIDERS = ("openai", "local") OPENAI_BASE_URL = "https://api.openai.com/v1" +POLL_ATTEMPTS = 4 +POLL_INTERVAL = 1.5 +MESSAGE_FORMAT_ERROR = "对话消息格式无效" +MESSAGE_ORDER_ERROR = "对话消息必须按用户描述、助手追问、用户补充的顺序提交" def require_full_admin(request): @@ -46,6 +52,101 @@ def resolve(org): return None +def request_response(cfg, payload, *, timeout=60.0, deadline=None): + """POST {base_url}/responses，并处理兼容服务异步返回的 in_progress。 + + deadline 为 time.monotonic() 的绝对时间预算；预算不足时抛 httpx.TimeoutException。 + HTTP 异常原样上抛，由调用方转换为业务错误。Authorization 仅在配置了 API Key 时发送。 + """ + base_url = cfg["base_url"].rstrip("/") + headers = {"Authorization": f"Bearer {cfg['api_key']}"} if cfg.get("api_key") else {} + + def budget_kwargs(): + if deadline is None: + return {} + remaining = deadline - time.monotonic() + if remaining <= 0: + raise httpx.TimeoutException("AI 请求超出时间预算") + return {"timeout": min(timeout, remaining)} + + def wait_before_poll(): + if deadline is not None and deadline - time.monotonic() <= POLL_INTERVAL: + raise httpx.TimeoutException("AI 请求超出时间预算") + time.sleep(POLL_INTERVAL) + + with httpx.Client(timeout=timeout) as client: + response = client.post(f"{base_url}/responses", headers=headers, json=payload, **budget_kwargs()) + response.raise_for_status() + result = response.json() + for _ in range(POLL_ATTEMPTS): + if not isinstance(result, dict) or result.get("status") != "in_progress": + break + response_id = result.get("id") + if not response_id: + break + wait_before_poll() + polled = client.get(f"{base_url}/responses/{response_id}", headers=headers, **budget_kwargs()) + polled.raise_for_status() + result = polled.json() + return result + + +def extract_output_text(result): + """校验响应已完成，返回首个非空 output_text（去除首尾空白）；否则 ValueError。""" + if not isinstance(result, dict) or result.get("status") != "completed": + raise ValueError("模型响应未完成") + output = result.get("output") + if not isinstance(output, list): + raise ValueError("模型响应缺少输出") + for item in output: + if not isinstance(item, dict) or not isinstance(item.get("content"), list): + continue + for part in item["content"]: + if not isinstance(part, dict) or part.get("type") != "output_text": + continue + text = part.get("text") + if isinstance(text, str) and text.strip(): + return text.strip() + raise ValueError("模型响应没有文本输出") + + +def validate_messages( + messages, + *, + max_messages=12, + user_chars=2000, + assistant_chars=4000, + total_chars=24000, + order_error=MESSAGE_ORDER_ERROR, +): + """统一校验对话消息：结构、长度、角色与 user/assistant 严格交替。""" + if not isinstance(messages, list) or not 1 <= len(messages) <= max_messages: + raise APIError(MESSAGE_FORMAT_ERROR) + normalized = [] + total = 0 + for item in messages: + if not isinstance(item, dict) or set(item) != {"role", "content"}: + raise APIError(MESSAGE_FORMAT_ERROR) + role, content = item["role"], item["content"] + if role not in ("user", "assistant"): + raise APIError(MESSAGE_FORMAT_ERROR) + if not isinstance(content, str) or not content.strip(): + raise APIError(MESSAGE_FORMAT_ERROR) + if len(content) > (user_chars if role == "user" else assistant_chars): + raise APIError(MESSAGE_FORMAT_ERROR) + total += len(content) + normalized.append({"role": role, "content": content}) + if total > total_chars: + raise APIError(MESSAGE_FORMAT_ERROR) + if ( + messages[0]["role"] != "user" + or messages[-1]["role"] != "user" + or any(left["role"] == right["role"] for left, right in zip(messages, messages[1:])) + ): + raise APIError(order_error) + return normalized + + def org_models_endpoint(org): """组织配置或环境回退的 base_url/api_key，用于探测模型列表。""" cfg = AIConfig.objects.filter(organization=org).first() diff --git a/finance/assistant.py b/finance/assistant.py index 17dc044..b8216da 100644 --- a/finance/assistant.py +++ b/finance/assistant.py @@ -2,49 +2,28 @@ import json import time import httpx +from django.core.serializers.json import DjangoJSONEncoder from django.http import JsonResponse -from .ai import resolve -from .analysis import METRICS, analyze, context, dimensions +from .ai import extract_output_text, request_response, resolve, validate_messages +from .analysis import METRICS, analyze, context, custom_columns, dimensions +from .formula_views import design_column from .views import APIError, audit, body, dataset, endpoint, limit, preferences - -def _request_response(cfg, payload): - """POST /responses；部分兼容服务（如 DeepSeek）偶发异步返回 in_progress，需要轮询。""" - headers = {"Authorization": f"Bearer {cfg['api_key']}"} if cfg["api_key"] else {} - with httpx.Client(timeout=60) as client: - response = client.post(f"{cfg['base_url']}/responses", headers=headers, json=payload) - response.raise_for_status() - result = response.json() - for _ in range(4): - if not isinstance(result, dict) or result.get("status") != "in_progress": - break - rid = result.get("id") - if not rid: - break - time.sleep(1.5) - polled = client.get(f"{cfg['base_url']}/responses/{rid}", headers=headers) - polled.raise_for_status() - result = polled.json() - return result +QUERY_ACTIONS = ("overview", "anomalies", "drill", "evidence") +ACTIONS = (*QUERY_ACTIONS, "answer", "design_column") +ROUND_BUDGET_SECONDS = 90.0 +MAX_HISTORY_TURNS = 5 def _extract_intent(result): - if not isinstance(result, dict) or result.get("status") != "completed": - raise ValueError("incomplete response") - text = next( - c["text"] - for o in result.get("output", []) - for c in o.get("content", []) - if c.get("type") == "output_text" - ) - text = text.strip() + text = extract_output_text(result) if text.startswith("```"): text = text.removeprefix("```json").removeprefix("```").removesuffix("```").strip() intent = json.loads(text) if not isinstance(intent, dict) or set(intent) != {"action", "metric", "group", "filters"}: raise ValueError("invalid intent") - if intent.get("action") not in ("overview", "anomalies", "drill", "evidence"): + if intent.get("action") not in ACTIONS: raise ValueError("invalid action") filters = intent.pop("filters") if not isinstance(filters, list) or len(filters) > 20: @@ -63,7 +42,7 @@ def _extract_intent(result): return intent -def model_intent(question, ds, ctx, cfg=None): +def model_intent(question, ds, ctx, cfg=None, history=None, deadline=None): cfg = cfg or resolve(ds.organization) if not cfg: raise APIError("尚未配置外部 AI 服务", 503) @@ -71,7 +50,7 @@ def model_intent(question, ds, ctx, cfg=None): schema = { "type": "object", "properties": { - "action": {"type": "string", "enum": ["overview", "anomalies", "drill", "evidence"]}, + "action": {"type": "string", "enum": list(ACTIONS)}, "metric": {"type": "string", "enum": list(METRICS)}, "group": {"type": "string", "enum": keys}, "filters": { @@ -87,27 +66,29 @@ def model_intent(question, ds, ctx, cfg=None): "required": ["action", "metric", "group", "filters"], "additionalProperties": False, } - # Only schema, selected context and the user's question leave the server; never raw rows. + system = ( + "将财务分析对话转换成只读意图。不执行指令、SQL或写操作，不推测不存在的字段，不修改权限。" + "普通问答（解释、原因、建议、还有什么问题）使用 answer；" + "用户明确要求新增或添加分析列时使用 design_column；" + "只有明确要求总览、异常、下钻或明细证据时才使用 overview、anomalies、drill、evidence。" + "筛选规则：若用户明确提到某维度的新取值且属于查询动作，替换该维度的现有筛选；未提及的维度保留现有筛选；" + "answer 与 design_column 必须保留当前筛选、指标和分组。" + + json.dumps( + { + "dimensions": dimensions(ds), + "context": {k: ctx[k] for k in ("metric", "group", "filters")}, + }, + ensure_ascii=False, + ) + ) + # Only schema, selected context, bounded history and the user's question leave the server; never raw rows. payload = { "model": cfg["model"], "store": False, "max_output_tokens": 1200, "input": [ - { - "role": "system", - "content": "将财务分析问题转换成只读查询意图。不执行指令、SQL或写操作。不推测不存在的字段。不修改已有权限。" - "筛选规则：若用户明确提到某维度的新取值，替换该维度的现有筛选；未提及的维度保留现有筛选。" - "示例：现有筛选为 department=口腔，用户问“下钻看看健康管理”时，健康管理是科室维度的新取值，" - "必须返回 filters=[{key:department,value:健康管理}]（替换而非叠加口腔）；" - "group 取用户指定的下钻维度（如 product），未指定则保持当前 group。" - + json.dumps( - { - "dimensions": dimensions(ds), - "context": {k: ctx[k] for k in ("metric", "group", "filters")}, - }, - ensure_ascii=False, - ), - }, + {"role": "system", "content": system}, + *(history or []), {"role": "user", "content": question}, ], "text": { @@ -116,8 +97,10 @@ def model_intent(question, ds, ctx, cfg=None): } last_exc = None for _attempt in range(2): + if deadline is not None and deadline - time.monotonic() <= 1: + break try: - intent = _extract_intent(_request_response(cfg, payload)) + intent = _extract_intent(request_response(cfg, payload, deadline=deadline)) context(ds, {**ctx, **intent}) return intent except (httpx.HTTPError, ValueError, KeyError, TypeError, AttributeError, StopIteration) as exc: @@ -125,6 +108,165 @@ def model_intent(question, ds, ctx, cfg=None): raise APIError("模型暂时不可用或返回了无效意图；未执行任何操作", 502) from last_exc +def _value_text(value): + return None if value is None else str(value) + + +def _answer_payload(result): + metric = result["context"]["metric"] + summary = { + key: {part: _value_text(value[part]) for part in ("current", "baseline", "delta", "change")} + for key, value in result["summary"].items() + } + rows = result["rows"][:5] + groups = { + "total": len(result["rows"]), + "truncated": len(result["rows"]) > 5, + "items": [ + { + "label": row["label"], + "filters": row["filters"], + "values": { + key: { + part: _value_text(value[part]) for part in ("current", "baseline", "delta", "change") + } + for key, value in row["values"].items() + }, + } + for row in rows + ], + } + columns = [column["name"] for column in result["context"].get("custom_columns", [])] + return { + "period": { + "start": result["context"]["start"], + "end": result["context"]["end"], + "comparison": result["context"]["comparison"], + "baseline_end": result["context"]["baseline_end"], + "current": result["context"]["current"], + "baseline": result["context"]["baseline"], + "missing": result["context"]["missing"], + }, + "filters": result["context"]["filters"], + "group": result["context"]["group"], + "metric": metric, + "metric_labels": METRICS, + "custom_columns": columns, + "summary": summary, + "comparable": result["comparable"], + "notice": result["notice"], + "anomaly_count": result["anomaly_count"], + "groups": groups, + } + + +def model_answer(question, ds, result, cfg, history=None, deadline=None): + payload_data = _answer_payload(result) + system = ( + "你是财务分析助手。只使用下面提供的服务端聚合数据回答用户当前问题，使用中文。" + "回答本次问题，不要复述整份数据；缺少期间、缺少字段或零基期时明确说明不可计算，不能填零或虚构数值。" + "相关性不等于因果；数据未支持的原因只能作为待核实假设，并明确标注为假设。" + "不得编造数据中没有的金额、比例、客户或期间；不得声称已修改数据或执行任何写操作。" + "聚合数据（JSON，金额为精确字符串）：" + + json.dumps(payload_data, ensure_ascii=False, cls=DjangoJSONEncoder) + ) + payload = { + "model": cfg["model"], + "store": False, + "max_output_tokens": 3000, + "input": [ + {"role": "system", "content": system}, + *(history or []), + {"role": "user", "content": question}, + ], + } + try: + result = request_response(cfg, payload, deadline=deadline) + if isinstance(result, dict) and result.get("status") == "incomplete": + reason = (result.get("incomplete_details") or {}).get("reason") + if reason == "max_output_tokens": + raise APIError("模型回答被截断（达到输出上限），请重试或缩短问题", 502) + raise APIError("模型回答不完整；未更改分析结果", 502) + text = extract_output_text(result).strip() + except APIError: + raise + except (httpx.HTTPError, ValueError, KeyError, TypeError, StopIteration) as exc: + raise APIError("模型暂时不可用或未返回有效回答；未更改分析结果", 502) from exc + if not text: + raise APIError("模型返回了空回答；未更改分析结果", 502) + return text + + +def _require_column_design(request): + if request.member.role not in ("admin", "analyst") or request.member.departments: + raise APIError("当前账号没有通过助手定义分析列的权限", 403) + + +def _column_context(ctx, columns): + return {**ctx, "custom_columns": columns} + + +def _column_design_reply(request, ds, ctx, columns, messages, cfg, provider, model): + _require_column_design(request) + design = design_column(messages, ds, ctx, cfg) + answer = ( + f"已整理列定义：{design['column']['name']}；请确认后添加到当前分析表。" + if design["complete"] + else "\n".join(design["questions"]) + ) + audit( + request, + "assistant.query", + ds.pk, + { + "provider": provider, + "model": model, + "action": "design_column", + "source": "model", + "kind": "column_design", + "stage": "completed", + "ok": True, + }, + ) + return JsonResponse( + { + "answer": answer, + "intent": { + "action": "design_column", + "metric": ctx["metric"], + "group": ctx["group"], + "filters": ctx["filters"], + }, + "context": _column_context(ctx, columns), + "version": ds.sha256, + "provider": provider, + "evidence": [], + "kind": "column_design", + "source": "model", + "model": model, + "design": design, + } + ) + + +def _rules_answer(result, metric): + value = result["summary"][metric] + if not result["comparable"]: + return "缺少期间 " + "、".join(result["context"]["missing"]) + "，不能给出比较或异常结论。" + if value["current"] is None: + return f"当前数据缺少计算{METRICS[metric]}所需的完整字段。" + change = value["change"] + change_text = f"{change:+.1%}" if change is not None else "不可计算（基期为零或数据缺失）" + answer = ( + f"当前范围{METRICS[metric]}为 {value['current']:,.2f}，较基期变化 {change_text}。" + f"按当前规则识别出 {result['anomaly_count']} 个异常分组。" + ) + candidates = [r for r in result["rows"] if r["anomaly"]][:3] + if candidates: + answer += " 优先核查：" + "、".join(r["label"] for r in candidates) + "。" + return answer + " 这是数据贡献分解，不是因果结论。" + + @endpoint(("POST",)) def assistant(request, dataset_id): ds = dataset(request, dataset_id, True) @@ -132,13 +274,36 @@ def assistant(request, dataset_id): question = data.get("question", "") if not isinstance(question, str) or not 1 <= len(question.strip()) <= 2000: raise APIError("问题需为 1 至 2,000 字") + mode = data.get("mode", "auto") + if mode not in ("auto", "column_design"): + raise APIError("对话模式无效") + history = data.get("history", []) + if not isinstance(history, list): + raise APIError("对话历史格式无效") + if len(history) > MAX_HISTORY_TURNS * 2: + raise APIError("对话最多保留 5 轮历史，请开始新对话") + conversation = validate_messages( + [*history, {"role": "user", "content": question.strip()}], + max_messages=(MAX_HISTORY_TURNS * 2) + 1, + user_chars=2000, + assistant_chars=4000, + total_chars=24000, + ) + history, current = conversation[:-1], conversation[-1] limit(f"assistant:{request.user.pk}", 20, 60) - ctx = context(ds, data.get("context", {})) + raw_context = data.get("context", {}) + ctx = context(ds, raw_context) + columns = custom_columns(ds, raw_context) pref, _ = preferences(request, ds) use_model = data.get("use_model", False) if type(use_model) is not bool: raise APIError("AI 模式必须为布尔值") - provider = "local" + if mode == "column_design": + _require_column_design(request) + if not use_model: + raise APIError("列设计需要启用 AI 模型", 400) + cfg = None + provider, model = "local", None if use_model: cfg = resolve(request.member.organization) if not cfg: @@ -146,8 +311,63 @@ def assistant(request, dataset_id): if data.get("consent") is not True: raise APIError("请先同意发送问题与当前筛选上下文", 400) provider = cfg["provider"] - intent = model_intent(question, ds, ctx, cfg) - else: + model = cfg["model"] + try: + if mode == "column_design": + return _column_design_reply(request, ds, ctx, columns, conversation, cfg, provider, model) + if use_model: + deadline = time.monotonic() + ROUND_BUDGET_SECONDS + intent = model_intent(current["content"], ds, ctx, cfg, history=history, deadline=deadline) + if intent["action"] == "design_column": + return _column_design_reply(request, ds, ctx, columns, conversation, cfg, provider, model) + if intent["action"] not in QUERY_ACTIONS: + intent = { + "action": "answer", + "metric": ctx["metric"], + "group": ctx["group"], + "filters": ctx["filters"], + } + result = analyze( + ds, + request.member, + {**ctx, **intent, "custom_columns": columns}, + pref, + ) + answer = model_answer(current["content"], ds, result, cfg, history=history, deadline=deadline) + audit( + request, + "assistant.query", + ds.pk, + { + "provider": provider, + "model": model, + "action": intent["action"], + "source": "model", + "kind": "answer", + "stage": "completed", + "ok": True, + }, + ) + return JsonResponse( + { + "answer": answer, + "intent": intent, + "context": result["context"], + "version": ds.sha256, + "provider": provider, + "evidence": [ + { + "label": r["label"], + "filters": r["filters"], + "values": r["values"][intent["metric"]], + } + for r in result["rows"][:5] + ], + "kind": "answer", + "source": "model", + "model": model, + } + ) action = ( "evidence" if "明细" in question or "证据" in question @@ -164,32 +384,51 @@ def assistant(request, dataset_id): for d in dimensions(ds): if d["name"] in question: intent["group"] = d["key"] - result = analyze(ds, request.member, {**ctx, **intent}, pref) - metric = intent["metric"] - value = result["summary"][metric] - if not result["comparable"]: - answer = "缺少期间 " + "、".join(result["context"]["missing"]) + "，不能给出比较或异常结论。" - elif value["current"] is None: - answer = f"当前数据缺少计算{METRICS[metric]}所需的完整字段。" - else: - change = value["change"] - change_text = f"{change:+.1%}" if change is not None else "不可计算（基期为零或数据缺失）" - answer = f"当前范围{METRICS[metric]}为 {value['current']:,.2f}，较基期变化 {change_text}。按当前规则识别出 {result['anomaly_count']} 个异常分组。" - candidates = [r for r in result["rows"] if r["anomaly"]][:3] - if candidates: - answer += " 优先核查：" + "、".join(r["label"] for r in candidates) + "。" - answer += " 这是数据贡献分解，不是因果结论。" - audit(request, "assistant.query", ds.pk, {"provider": provider, "intent": intent}) - return JsonResponse( - { - "answer": answer, - "intent": intent, - "context": result["context"], - "version": ds.sha256, - "provider": provider, - "evidence": [ - {"label": r["label"], "filters": r["filters"], "values": r["values"][metric]} - for r in result["rows"][:5] - ], - } - ) + result = analyze(ds, request.member, {**ctx, **intent, "custom_columns": columns}, pref) + answer = _rules_answer(result, intent["metric"]) + audit( + request, + "assistant.query", + ds.pk, + { + "provider": provider, + "model": None, + "action": intent["action"], + "source": "rules", + "kind": "answer", + }, + ) + return JsonResponse( + { + "answer": answer, + "intent": intent, + "context": result["context"], + "version": ds.sha256, + "provider": provider, + "evidence": [ + {"label": r["label"], "filters": r["filters"], "values": r["values"][intent["metric"]]} + for r in result["rows"][:5] + ], + "kind": "answer", + "source": "rules", + "model": None, + } + ) + except APIError as exc: + if use_model: + audit( + request, + "assistant.query", + ds.pk, + { + "provider": provider, + "model": model, + "action": "unknown", + "source": "model", + "kind": "error", + "stage": "model", + "ok": False, + "status": exc.status, + }, + ) + raise diff --git a/finance/formula_views.py b/finance/formula_views.py index 5e7d78d..f2b05a8 100644 --- a/finance/formula_views.py +++ b/finance/formula_views.py @@ -8,32 +8,18 @@ from django.db.models import Q from django.http import FileResponse, JsonResponse from django.utils import timezone +from .ai import extract_output_text, request_response, resolve, validate_messages from .analysis import METRICS, context, dimensions from .formula_columns import available_metrics, ensure_available_metrics, validate_column from .models import Dataset, FormulaDefinition, FormulaExport from .views import APIError, audit, body, dataset, endpoint, limit -def design_column(messages, ds, ctx): - if not (settings.AI_ENABLED and settings.OPENAI_API_KEY and settings.OPENAI_MODEL): +def design_column(messages, ds, ctx, cfg=None): + cfg = cfg or resolve(ds.organization) + if not cfg: raise APIError("尚未配置外部 AI 服务", 503) - if not isinstance(messages, list) or not 1 <= len(messages) <= 12: - raise APIError("对话消息格式无效") - if any( - not isinstance(item, dict) - or set(item) != {"role", "content"} - or item["role"] not in ("user", "assistant") - or not isinstance(item["content"], str) - or len(item["content"]) > 1000 - for item in messages - ): - raise APIError("对话消息内容无效") - if ( - messages[0]["role"] != "user" - or messages[-1]["role"] != "user" - or any(left["role"] == right["role"] for left, right in zip(messages, messages[1:])) - ): - raise APIError("对话消息必须按用户描述、助手追问、用户补充的顺序提交") + messages = validate_messages(messages) available = available_metrics(ds) @@ -117,9 +103,9 @@ def design_column(messages, ds, ctx): ) ) payload = { - "model": settings.OPENAI_MODEL, + "model": cfg["model"], "store": False, - "max_output_tokens": 1400, + "max_output_tokens": 2000, "input": [{"role": "system", "content": system}, *messages], "text": { "format": { @@ -131,17 +117,9 @@ def design_column(messages, ds, ctx): }, } try: - with httpx.Client(timeout=30) as client: - response = client.post( - "https://api.openai.com/v1/responses", - headers={"Authorization": f"Bearer {settings.OPENAI_API_KEY}"}, - json=payload, - ) - response.raise_for_status() - result = response.json() - if not isinstance(result, dict) or result.get("status") != "completed": - raise ValueError("incomplete response") - text = next(c["text"] for o in result["output"] for c in o["content"] if c["type"] == "output_text") + text = extract_output_text(request_response(cfg, payload, timeout=30)) + if text.startswith("```"): + text = text.removeprefix("```json").removeprefix("```").removesuffix("```").strip() design = json.loads(text) if ( set(design) != {"complete", "questions", "column"} diff --git a/finance/views.py b/finance/views.py index 49dcf8d..f582319 100644 --- a/finance/views.py +++ b/finance/views.py @@ -26,7 +26,7 @@ from openpyxl.cell import WriteOnlyCell from .analysis import METRICS, analyze, context, custom_columns, default_preferences, dimensions, scoped_facts from .importing import file_digest, profile_file, suggest_mapping, validate_mapping -from .models import AIConfig, Audit, Dataset, Investigation, Membership, Preference, Throttle +from .models import Audit, Dataset, Investigation, Membership, Preference, Throttle class APIError(Exception): @@ -133,19 +133,24 @@ def serialize_dataset(ds, detailed=False, restricted=False, raw_access=True): @endpoint(public=True) def session(request): ai_enabled = settings.AI_ENABLED and bool(settings.OPENAI_API_KEY and settings.OPENAI_MODEL) + ai_model = None + ai_provider = None if request.user.is_authenticated: try: - cfg = AIConfig.objects.filter(organization=request.user.membership.organization).first() + from .ai import resolve + + cfg = resolve(request.user.membership.organization) + ai_enabled = bool(cfg) if cfg: - ai_enabled = ( - cfg.enabled and bool(cfg.model) and (cfg.provider == "local" or bool(cfg.api_key)) - ) + ai_model, ai_provider = cfg["model"], cfg["provider"] except Membership.DoesNotExist: pass data = { "csrf": get_token(request), "user": None, "ai_enabled": ai_enabled, + "ai_model": ai_model, + "ai_provider": ai_provider, } if request.user.is_authenticated: try: diff --git a/frontend/src/App.tsx b/frontend/src/App.tsx index 5cbc3d4..63bfe41 100644 --- a/frontend/src/App.tsx +++ b/frontend/src/App.tsx @@ -48,6 +48,7 @@ import { labels, roleLabels, type Analysis, + type AssistantReply, type Context, type Dataset, type FormulaColumn, @@ -514,6 +515,9 @@ export function App() { if (!session.user) return ; const user = session.user; const canUpload = user.role !== "viewer" && !user.departments.length; + const canDesign = + (user.role === "admin" || user.role === "analyst") && + !user.departments.length; return (