🚀 Quick Start

This system is a privately deployed enterprise knowledge assistant. After administrators connect documents and build an index, users can ask questions in natural language. Answers are generated from indexed content and include cited sources.

🧭 Concepts

Knowledge Index

The system reads documents, splits them into searchable chunks, embeds them, and stores vectors in ChromaDB. At query time it retrieves relevant chunks before asking the model to answer.

Cited Sources

Answers show cited files, pages, or snippets so users can verify the basis of the response.

☁️ Cloud Storage

Open Cloud Storage, select a provider, and follow the setup guide in the dialog. OAuth providers require a callback URL in the provider developer console. Access Key providers should use a read-only sub-account.

• OAuth: Google Drive, OneDrive, Dropbox.
• Access Key: AWS S3, Alibaba Cloud OSS, Tencent Cloud COS.
• Password: SFTP.

After adding a remote, click Mount. The mount path is /mnt/<remote-name>. You can then configure sync interval and maximum index file size.

If Google OAuth is in testing mode, refresh tokens may expire after 7 days. For production use, publish the app or use Reconnect in Cloud Storage.

📚 Indexing

Add local directories, mounted cloud directories, or website URLs in Index Management. Image OCR is off by default; enable it only when you really need text extraction from images.

If indexing fails or is interrupted, failed batch data is cleaned up to avoid broken index records affecting queries.

🗂️ Document Organization and Naming

Good folder structure and file naming improves scan speed, reduces duplicate indexing, and makes source citations clearer.

/mnt/knowledge/
  products/
    drone/
    camera/
  policies/
  training/
  faq/
  archive/

• Include product, version, region, or year in file names, such as Matrice_4_Specs_v2026_EN.pdf.
• Keep one topic per folder. Avoid mixing unrelated files in one large directory.
• Put historical versions into archive instead of mixing them with the latest documents.
• Prefer text-based PDFs over scanned PDFs to improve indexing speed and answer quality.
• Split very large files by chapter and keep files under the maximum index size configured in Cloud Storage.

🌐 Website Index

Website indexing brings public pages or internal documentation sites into the knowledge base. Keep same-domain crawling enabled and limit maximum pages for predictable runtime.

🤖 Models

Administrators can configure DeepSeek, OpenAI, Gemini, Qwen, Doubao, or local Ollama. API keys are centrally managed by administrators. Users only select from available models.

💬 Chat

• Ask a question directly to search the knowledge base.
• Expand Sources under an answer to verify the basis.
• Uploaded files are used as context for the current question.
• Administrators can return to the admin console from chat. Advanced and Standard users only see the features they need.

📋 Smart Response

Advanced users and administrators can upload bids, requirement sheets, and questionnaires. The system extracts requirements, answers them item by item, and exports results to Excel.

👥 Users and Roles

💬 Lark and WeCom Integration

The system can connect IM app messages to the same knowledge-base Q&A flow. Users send text messages to the bot in Lark or WeCom, and the system creates a Standard user automatically, keeps conversation context, and replies in the IM app.

The IM integration currently focuses on text Q&A. File upload, smart response, and model preference workflows should still use the web app.

Lark / Feishu Setup

1. Open Lark Open Platform or Feishu Open Platform and create an internal enterprise app.
2. In Credentials & Basic Info, record App ID and App Secret. URL verification itself does not require them, but the system needs them to send replies and fetch user information.
3. In Event Subscription, set the request URL to https://your-domain/api/lark/webhook.
4. Configure the Verification Token. If Encrypt Key is enabled, also set it in the server deployment configuration. In the current code, Verification Token and Encrypt Key are optional: if empty, token validation is skipped and events are treated as plain JSON.
5. Subscribe to the message receive event: im.message.receive_v1.
6. Enable the required permissions, then publish a new app version.

The Lark callback URL is https://your-domain/api/lark/webhook. Initial URL verification only needs this endpoint to be reachable. To let the bot reply to users, add LARK_APP_ID and LARK_APP_SECRET to the server deployment configuration, such as /opt/consultant-linux/.env or the app environment section in docker-compose.yml. Restart the app container after changes. LARK_VERIFY_TOKEN and LARK_ENCRYPT_KEY are only needed if the corresponding validation/encryption is enabled in Lark.

LARK_APP_ID=
LARK_APP_SECRET=
LARK_VERIFY_TOKEN=
LARK_ENCRYPT_KEY=

WeCom Setup

1. Open the WeCom admin console, go to App Management, and create a custom app.
2. Record CorpID, AgentId, and app Secret.
3. Open the app's message receiving / API message receiving settings.
4. Set URL to https://your-domain/api/wecom/webhook.
5. Set Token and EncodingAESKey, and use the same values in the server deployment configuration.
6. When saving, WeCom sends a URL verification request. The system decrypts and returns echostr automatically.
7. Set the app visible scope to the departments or members who should use Q&A.

These settings are also deployment-level settings, not admin UI fields. Add them to /opt/consultant-linux/.env or the app environment section in docker-compose.yml, then restart the app container.

WECOM_CORP_ID=
WECOM_AGENT_ID=
WECOM_SECRET=
WECOM_TOKEN=
WECOM_ENCODING_AES_KEY=

IM users are created as Standard users on first question. Administrators can upgrade them to Advanced or Admin in User Management.

🎨 Branding and Notices

Branding lets you set the tool name, company name, logo, bot avatar, and favicon. The tool name is user-defined content and is not translated by language switching.

System Settings can configure SMTP email and Aliyun SMS for verification codes.

💾 Backup

Back up PostgreSQL data, ChromaDB data, rclone configuration, uploaded assets, and /data/rag_config.json. Restore data before starting containers on a new server.