v0.1.0 — 首个可运行版本 / First runnable release

简体中文 · English

容器镜像 / Container image —— 公开包，无需登录 / public package, no login required

docker pull ghcr.io/groundedcore/nexusdesk-quickstart:latest
docker pull ghcr.io/groundedcore/nexusdesk-quickstart:ae97956bdc7396c734030f40de457518ee42b597

源码归档见本页底部附件。Source archives are attached to this release.

---

简体中文

v0.1.0 — 2026-10-09

首个可运行版本，客服核心流程已完整可用。

新增

平台能力

• Agent 配置与执行：配置并发布 Agent，通过异步 LangGraph ReAct 循环执行推理与工具调用。
• 知识库与检索：接入文档、配置分片、预览内容、构建向量索引并测试检索效果。
• 统一模型接入：管理 Chat、Embedding、Rerank、OCR 与语音模型的连接、目录及版本方案。
• 业务工具集成：接入业务 API，管理调用凭据、参数与执行权限。
• 人工协同与工单：支持会话接管、工单拟定与确认。
• 渠道与应用接入：通过渠道接口和应用 API 接入业务系统，含嵌入聊天与企业身份登录（OIDC、企业微信、钉钉、飞书）。
• 运行观测与评估：查看执行记录、工具调用与评估结果。

共 13 个领域模块、29 个数据库迁移；控制台支持简体中文、繁体中文、英文、印地语。

部署拓扑

• quickstart：App + PostgreSQL 两个容器，无需模型密钥，内置演示模型。
• production：Web、API、Runtime Worker、Knowledge Worker 四个常驻服务，外加一次性迁移任务；中间件复用外部 PostgreSQL、Milvus 与 MinIO，不重复启动。

预置数据

• 示例客服：知识库、模型方案（已发布）、Agent（已发布 v1）与 sample-support 会话，开箱即可对话。
• 9 个行业案例：9 个知识库、18 个 Agent 草稿、36 篇标注为虚构的案例文档，资源名带 [案例] 前缀。
• 案例 Agent 的提示词按「服务准则 + 表达方式 + 行业语气」三层组织：行为约束与话术分离，模型不会把护栏当谈资念给用户。

体验模式 HTTPS

• 体验模式自带私有 CA 与服务器证书，容器启动时生成，因此通过局域网 IP 或内网域名访问时同样处于安全上下文，crypto.randomUUID、navigator.clipboard 等仅限安全上下文的 API 可正常使用。
• 证书在主机列表变化或临近过期时自动重签，CA 保持不变，已导入的 ca.crt 无需重新导入。

已知限制

• 体验模式使用模拟模型，不代表真实回答质量；不连接真实订单业务，不提供真实 OCR、向量检索或 GPU 推理。
• 体验模式会为未携带身份的请求注入 admin 角色令牌，不要暴露到公网；确需放开绑定时用防火墙限制来源。
• 生产拓扑不自带 TLS，需由宿主机或入口代理终止。
• 真实 Parser / Milvus / S3 尚未完成联调，详见模块交付说明。
• 项目处于持续开发阶段，尚未达到生产可用。

升级说明

首个版本，无升级路径。数据库迁移随容器启动自动执行，且不支持回滚——回滚镜像不会回滚数据库结构，因此迁移必须保持向前兼容（加列可空或有默认值、删列分两步）。详见数据库迁移。

---

English

v0.1.0 — 2026-10-09

First runnable release. The core customer-service flow is complete and usable.

Added

Platform capabilities

• Agent configuration and execution: configure and publish agents, then run reasoning and tool calls through an asynchronous LangGraph ReAct loop.
• Knowledge base and retrieval: ingest documents, configure chunking, preview content, build vector indexes, and test retrieval quality.
• Unified model access: manage connections, catalogs, and versioned profiles for Chat, Embedding, Rerank, OCR, and speech models.
• Business tool integration: connect business APIs and manage call credentials, parameters, and execution permissions.
• Human collaboration and tickets: conversation takeover plus ticket drafting and confirmation.
• Channels and application access: connect through channel endpoints and application APIs, including the embedded chat widget and enterprise identity login (OIDC, WeCom, DingTalk, Feishu).
• Observability and evaluation: review execution records, tool calls, and evaluation results.

13 domain modules and 29 database migrations. The console ships in Simplified Chinese, Traditional Chinese, English, and Hindi.

Deployment topologies

• quickstart: two containers (app plus PostgreSQL), no model key required, ships with a mock model.
• production: four long-running services (web, API, runtime worker, knowledge worker) plus a one-shot migration task. Middleware is reused from an existing PostgreSQL, Milvus, and MinIO rather than started again.

Seeded data

• Sample support setup: knowledge base, model profile (published), agent (published v1), and a sample-support conversation, ready to chat with out of the box.
• Nine industry case sets: 9 knowledge bases, 18 agent drafts, and 36 documents, each marked as fictional and prefixed with [案例].
• Case agent prompts are organised in three layers: conduct rules, writing style, and per-industry voice. Behaviour constraints stay separate from phrasing, so the model does not recite its own guardrails back to the user.

HTTPS for the quickstart demo

• The quickstart bundle ships its own private CA and server certificate, generated when the container starts. Reaching the demo over a LAN IP or an internal hostname is therefore still a secure context, so secure-only APIs such as crypto.randomUUID and navigator.clipboard behave normally.
• The leaf certificate is re-issued automatically when the host list changes or it nears expiry, while the CA stays the same, so an already imported ca.crt never needs to be imported again.

Known limitations

• The demo uses a mock model and does not represent real answer quality. It does not connect to real order systems and provides no real OCR, vector search, or GPU inference.
• The demo injects an admin-role token for requests that carry no identity. Do not expose it to the public internet; if you must widen the bind address, restrict the source with a firewall.
• The production topology does not terminate TLS itself; the host or an ingress proxy is expected to.
• Real parser, Milvus, and S3 integrations have not been validated end to end. See module delivery notes.
• The project is under active development and is not production ready.

Upgrade notes

This is the first release, so there is no upgrade path. Database migrations run automatically when the container starts and cannot be rolled back — reverting the image does not revert the schema, so migrations must stay forward compatible (new columns nullable or defaulted, dropped columns removed over two releases). See database migrations.