审脉 AuditPilot

面向审计交付场景的企业级 AI Agent 工作台：把证据检索、工具执行、Harness 评测、人工复核与整改闭环放进同一条可追溯工作流。

Why AuditPilot

很多 Agent 项目停留在聊天框或 Demo。AuditPilot 选择一个更“硬”的落地场景：企业审计交付。它需要证据、控制、风险、复核、报告和整改闭环，也天然要求可追溯、可回归、可解释。

AuditPilot 的目标不是替代审计师，而是把审计师反复执行的取证、映射、检查、补证和交付动作，组织成一套可治理的 Agent 工作流。

解决什么业务问题

项目不在没有客户实测的情况下宣称固定的“节省工时”或“准确率”。业务价值应通过取证周期、底稿一次复核通过率、证据充分率和整改按期关闭率在试点项目中持续度量。

审计场景覆盖

当前内置 6 个可直接立项的高频数字化审计模板，合计覆盖 4 类标准、34 个控制主题、36 类证据输入和 25 类交付物：

这些数字来自 services/audit_templates.py 中版本化模板的去重统计，代表开箱范围，不代表已覆盖所有行业和全部审计业务。企业可接入自有标准库、控制库和证据源扩展场景。

What it does

Screenshots

Audit workspace	Agent runtime

Agent collaboration	Layered evaluation

Architecture

Audit request
  -> Hybrid Intent Router
  -> Working + Episodic + Profile Memory
  -> Planner / Evidence / Control / Risk / Compliance / Remediation / Verification / Delivery
  -> Bounded dependency-aware Agent Loop
  -> Agentic RAG + Evidence Graph + Skills / MCP-style Tools
  -> Safety Gate + Reflection + Human Review
  -> 9-layer Component Evaluation + Release Gate
  -> Delivery Package + Governed Experience Candidate

Design boundaries:

• LLMs help with understanding, summarization and explanation.
• Evidence gaps, quality gates, permissions, risk signals and delivery state stay auditable.
• High-risk or low-confidence outputs are routed to evidence completion and human review.

快速运行

要求 Python 3.10 或更高版本。项目默认支持无模型密钥运行：未配置大模型时会进入确定性回退模式，审计工作流、RAG、评测和界面仍可体验。

Windows PowerShell

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
Copy-Item config.env.example config.env
python start.py

macOS / Linux

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp config.env.example config.env
python start.py

启动完成后，按终端输出的地址在浏览器中打开应用。停止服务时在终端按 Ctrl+C。

基本操作

1. 在「Agent 协作」中输入审计对象、标准或风险场景，观察多 Agent 的规划、证据、控制、风险与整改协作。
2. 在「审计项目」中建立审计范围，运行控制测试、Deep Research、证据请求和整改闭环。
3. 在「知识库」中维护可检索的审计知识；首次运行会从公开种子知识自动建立本地 RAG 存储。
4. 在「Agent 评测」中运行分层评测，查看任务、轨迹、工具、证据、安全、上下文与鲁棒性结果，以及发布门禁。
5. 在「Skills / MCP」中检查工具 Schema、权限声明、调用记录和运行状态。

可选配置

复制 config.env.example 后，只在本地 config.env 中填写所需配置：

默认安装不包含 Torch / Transformers。只有在确实需要本地语义向量时，才安装 requirements-embeddings.txt 并启用 RAG_ENABLE_EMBEDDINGS=1；默认混合检索不依赖本地大模型。

Security

Do not commit real API keys.

• Put local secrets in config.env.
• Use platform secrets for deployment.
• config.env, .env*, runtime data, logs, model artifacts and local databases are ignored by Git.
• Example config files use placeholders only.
• SECURITY_MODE=local is restricted to a loopback bind. Any shared or network deployment must use SECURITY_MODE=enforced, secret-managed bearer tokens and an audit-log signing key.
• Permissions are checked at the HTTP and Skill boundaries; repositories additionally enforce tenant and project membership.
• Uploads are streamed with a hard size limit, content/extension checks, hashes, credential scanning and prompt-injection quarantine.
• Liveness and dependency-aware readiness are exposed separately; readiness fails when security or persistent stores are not usable.

Public repository scope

公开仓库只保留可运行产品代码、必要的公开种子数据、回归测试、展示截图、README 与启动配置。以下内容不会进入 Git：

• API 密钥、数据库密码和个人配置；
• 本地运行记录、会话记忆、评测结果、上传文件和生成的 RAG 存储；
• 求职 JD、面经、项目内部设计文档和完整归档；
• 本地模型、IDE 配置、日志及临时生成文件。

Validation

.\.venv\Scripts\python.exe -m compileall -q agents services rag web tests scripts
.\.venv\Scripts\python.exe -m pytest -q
.\.venv\Scripts\python.exe -m pytest --cov=services.record_store --cov=services.security --cov=services.agent_runtime --cov=services.conversation_memory --cov=rag.agentic_rag --cov-fail-under=75
.\.venv\Scripts\python.exe scripts\audit_repro.py
ruff check agents services rag web tests
bandit -q -ll -r agents services rag web -x tests
pip-audit -r requirements.txt

Current tests cover authentication/RBAC/ABAC, tenant and project isolation, optimistic locking, signed audit checkpoints, upload injection blocking, intent routing, memory compaction, Skill input/output contracts, adaptive multi-role runtime execution, immutable-artifact delivery verification, fail-closed RAG filtering, evidence lineage, governed experience reuse, independent-dataset evaluation semantics, repository-bound Harness gating and every visible product endpoint.

Project layout

agents/                 audit agent chain and control library
services/               runtime, memory, router, skills, safety, evaluation, delivery
rag/                    agentic RAG and local knowledge-store runtime
knowledge_graph/        optional graph construction and Neo4j adapter
training/               evaluation and offline training entry points
web/                    FastAPI application and APIs
templates/ + static/    product UI
tests/                  regression tests
docs/screenshots/       public product screenshots only
data/seed_knowledge/    public starter knowledge; runtime data is ignored

Reality notes

• Runtime records, tool calls, evaluations, memory and audit cases are generated by real code and persisted locally.
• Audit, evaluation, Agent task/event and conversation-memory records use tenant-scoped SQLite WAL storage with optimistic concurrency; RAG chunks remain a single-node snapshot/index in this release.
• The built-in evaluation suite is a regression/smoke baseline, not proof of business accuracy. Release-grade claims require an independent held-out dataset and domain-expert review.
• The included bearer-token RBAC is a deployable baseline, not a replacement for enterprise OIDC/SSO, external KMS, object storage or a multi-node database.
• Role collaboration is an observable, contract-bound orchestration inside one runtime process; it is not presented as a fleet of independently deployed autonomous services.
• Untested throughput, accuracy or availability claims are intentionally not presented as project facts.
• SFT / RLHF-style model training is treated as an offline extension; the Web app focuses on workflow, evaluation and data preparation.