SlideRAG: PPT-Centric Multimodal RAG for Study Preview & Exam Review

English | 简体中文

🆕 Updates

• [04/2026] 🔥 SlideRAG is now open-source.
• [04/2026] ✨ Added Feishu and WhatsApp channels to bring the same SlideRAG agent to more chat platforms.

🚀 SlideRAG

🎓 SlideRAG is an end-to-end assistant for understanding PPT/PPTX files as multimodal learning materials.

🧠 Unlike text-only QA systems, SlideRAG treats each slide as a structured multimodal unit and combines parsing, retrieval, and agent tool-calling.

📌 It is designed for two key learning scenarios: before-class preview and before-exam review.

🎬 Demo Showcase

This video presents SlideRAG from a product perspective and walks through the full user flow: upload slides, inspect parsing results, and run multi-turn QA over PPT content.

283d7dc3010f5cc719306119e71a9e2e.mp4

Here is an example that demonstrates the performance of SlideRAG in understanding the content and structure of multimodal slides:

QA case snapshots:

✨ Key Features

• 🖼️ PPT-first multimodal RAG pipeline: Uses a unified multimodal parser and a graph-and-vector hybrid retrieval engine to support grounded QA across text, images, tables, and equations.
• 🪄 Hidden-information expansion for concise slides: Detects high-compression pages and expands implicit content into grounded explanatory text.
• 🔗 Page-topic extraction and structural linking: Extracts per-page topics and links related slides to model section-level continuity in long decks.
• 🤝 Easy to use: One backend supports Web, QQ, Feishu, WeChat, and WhatsApp bridge mode, making the assistant accessible in familiar study workflows.

🧩 Framework

SlideRAG follows a retrieval-augmented agent workflow: 1. Parse PPT/PPTX into typed multimodal items with page metadata. 2. Perform PPT-oriented enhancement (hidden-info expansion + topic extraction/linking). 3. Build unified multimodal knowledge storage for hybrid retrieval. 4. Use tool-calling agent loop to retrieve evidence and trigger optional image understanding. 5. Return grounded answers through Web/QQ/Feishu/WeChat/WhatsApp channels.

🚀 Quick Start

This section helps you run SlideRAG quickly for web usage, then optionally connect it to QQ, Feishu, WeChat, or WhatsApp.

1. Clone and install

git clone https://github.com/Hitlh/SlideRAG.git
cd SlideRAG

# Core dependencies
pip install -e .

# Optional channel dependencies
pip install -e .[qq]
pip install -e .[feishu]
pip install -e .[weixin]
pip install -e .[whatsapp]
pip install -e .[channels]

Because this project focuses on PPT understanding, install LibreOffice as an extra system dependency:

• Ubuntu/Debian: sudo apt-get install libreoffice
• Windows: download installer from the official website: https://www.libreoffice.org/
• macOS: brew install --cask libreoffice

2. Configure environment variables

Create a .env file by copying content from env.project.example, then fill in your keys and model settings.

Note: If you want to tune advanced parser/context behavior (for example, SUMMARY_LANGUAGE, hidden-expansion options, and context window settings), edit the Advanced parser/context options (optional) section in env.project.example.

Tip for users in mainland China: MinerU uses Hugging Face by default. If Hugging Face access is unstable, switch to ModelScope before running:

export MINERU_MODEL_SOURCE=modelscope

2.1 API keys and base models

OPENAI_API_KEY=your_openai_api_key
OPENAI_BASE_URL=your_base_url

# Text and vision models used by SlideRAG pipeline
TEXT_LLM_MODEL=gpt-5.4
VLM_MODEL=gpt-5.4

2.2 Agent model provider

# Agent provider for rag_agent loop: openai | anthropic
AGENT_PROVIDER=openai
AGENT_MODEL=gpt-5.4

# Anthropic provider settings (only required when AGENT_PROVIDER=anthropic)
# If empty, runtime may fall back to OPENAI_API_KEY / OPENAI_BASE_URL.
ANTHROPIC_API_KEY=
ANTHROPIC_BASE_URL=

3. Start the web app

streamlit run client/app.py

After startup, open the Streamlit URL shown in terminal and start asking questions about your PPT files.

Chat App Integration

SlideRAG can run the same QA agent through QQ, Feishu, WeChat, and WhatsApp bridge mode.

QQ setup (requires QQ extras)

1. Create a QQ bot.
2. Go to the QQ Open Platform (https://q.qq.com/#/), sign in, and create your bot.
3. In your bot console, open "开发控制" and copy APPID and APPSecret.
4. Go to "沙箱配置", then in "在消息列表配置" choose "添加成员", enter your QQ number, and scan the QR code.
5. Put the files you want to chat with in a folder (default: ./uploaded_docs).
6. Configure QQ environment variables:

QQ_ENABLED=false                    # Set to true to enable QQ integration
QQ_APP_ID=                          # QQ bot APPID from the Open Platform
QQ_SECRET=                          # QQ bot APPSecret from the Open Platform
QQ_ALLOW_FROM=*                     # Allowed senders; use * to accept all
QQ_TARGET_FILE=                     # Default file to chat with; can switch via /file <filename>
QQ_UPLOADED_DOCS_DIR=./uploaded_docs # Directory that stores your source files

# Startup ready notification
QQ_STARTUP_NOTIFY_ENABLED=true      # Send a startup message when agent is ready
QQ_STARTUP_NOTIFY_MESSAGE=rag agent is ready. # Startup message content
QQ_STARTUP_NOTIFY_CHAT_ID=          # Target chat/user ID for startup notification

7. Run QQ backend:

python3 -m client.qq.runtime
# or
sliderag-qq

8. Start chatting. You can switch the active document with: /file <filename>.

Feishu setup (requires Feishu extras)

1. Go to the Feishu/Lark open platform. China users can use open.feishu.cn, and global users can use open.larksuite.com.
2. Create an app and add the Bot capability.
3. In Developer Configuration > Permissions, enable the permissions needed for messaging:

• im:message
• im:message.p2p_msg:readonly

4. In Developer Configuration > Events and Callbacks, use long connection for event delivery, then add the event:

• im.message.receive_v1

5. Save the App ID and App Secret in Credentials & Basic Information.
6. Publish the app.
7. Put the files you want to chat with in a folder (default: ./uploaded_docs).
8. Configure Feishu environment variables:

FEISHU_ENABLED=true
FEISHU_APP_ID=
FEISHU_APP_SECRET=
FEISHU_ENCRYPT_KEY=
FEISHU_VERIFICATION_TOKEN=
FEISHU_ALLOW_FROM=*
# feishu (China) | lark (global)
FEISHU_DOMAIN=feishu

# Optional startup target file for auto-ingest
FEISHU_TARGET_FILE=

# Runtime directories
FEISHU_UPLOADED_DOCS_DIR=./uploaded_docs
FEISHU_INGEST_OUTPUT_DIR=./output
FEISHU_RAG_WORKING_DIR=./rag_storage_by_feishu_file
FEISHU_RUNTIME_STATE_DIR=./rag_storage_feishu_runtime

# Startup ready notification
FEISHU_STARTUP_NOTIFY_ENABLED=true
FEISHU_STARTUP_NOTIFY_MESSAGE=agent is ready.
FEISHU_STARTUP_NOTIFY_CHAT_ID=

9. Run Feishu backend:

python3 -m client.feishu.runtime

10. Send one message in Feishu first if you want to use FEISHU_ALLOW_FROM or FEISHU_STARTUP_NOTIFY_CHAT_ID, then check runtime logs for the sender_id value.

WeChat setup (requires WeChat extras)

1. Put the files you want to chat with in a folder (default: ./uploaded_docs).
2. Configure WeChat environment variables:

WEIXIN_ENABLED=true                # Set to true to enable WeChat integration
WEIXIN_ALLOW_FROM=*                # Allowed senders; use * to accept all
WEIXIN_TARGET_FILE=                # Default file to chat with; can switch via /file <filename>
WEIXIN_UPLOADED_DOCS_DIR=./uploaded_docs # Directory that stores your source files
WEIXIN_STARTUP_NOTIFY_ENABLED=true # Send a startup message when agent is ready
WEIXIN_STARTUP_NOTIFY_MESSAGE=agent is ready. # Startup message content
WEIXIN_STARTUP_NOTIFY_CHAT_ID=     # Target chat/user ID for startup notification

3. Run WeChat backend:

python3 -m client.weixin.runtime
# use -r to force re-login
python3 -m client.weixin.runtime -r

4. Scan QR code on first login.
5. Start chatting. You can switch the active document with: /file <filename>.

WhatsApp setup (requires WhatsApp extras + local bridge)

SlideRAG uses a local WebSocket bridge with protocol messages (message/status/qr/error, send/send_media).

1. Start the bundled local WhatsApp bridge in this repository (default endpoint: ws://127.0.0.1:3001) and keep it running. Example bridge startup:

cd /path/to/SlideRAG/bridge
npm install
npm run build

export BRIDGE_PORT=3001
export BRIDGE_TOKEN=replace_with_secret
export AUTH_DIR=/abs/path/to/wa_auth
npm start

BRIDGE_TOKEN must be the same value as WHATSAPP_BRIDGE_TOKEN in SlideRAG .env.

Tip: If you test by sending messages from the same WhatsApp account (self-message), start bridge with export WHATSAPP_ACCEPT_FROM_ME=true.

2. Configure WhatsApp environment variables:

WHATSAPP_ENABLED=true
WHATSAPP_ALLOW_FROM=*                      # Allowed sender/chat ids; use * to accept all
WHATSAPP_BRIDGE_URL=ws://127.0.0.1:3001    # Local bridge endpoint
WHATSAPP_BRIDGE_TOKEN=replace_with_secret  # Must match bridge auth token
WHATSAPP_RECONNECT_DELAY_S=5
WHATSAPP_SEND_RETRY_ATTEMPTS=3             # Outbound send retry attempts
WHATSAPP_SEND_RETRY_DELAY_MS=400           # Outbound retry delay in milliseconds
WHATSAPP_ACCEPT_GROUP_MESSAGES=true        # Accept group inbound messages
WHATSAPP_REQUIRE_MENTION_IN_GROUP=true     # In group chats, reply only when bot is mentioned
WHATSAPP_TARGET_FILE=                      # Optional; can switch via /file <filename>
WHATSAPP_UPLOADED_DOCS_DIR=./uploaded_docs
WHATSAPP_INGEST_OUTPUT_DIR=./output
WHATSAPP_RAG_WORKING_DIR=./rag_storage_by_whatsapp_file
WHATSAPP_RUNTIME_STATE_DIR=./rag_storage_whatsapp_runtime
WHATSAPP_STARTUP_NOTIFY_ENABLED=true
WHATSAPP_STARTUP_NOTIFY_MESSAGE=agent is ready.
WHATSAPP_STARTUP_NOTIFY_CHAT_ID=

3. Run WhatsApp backend:

python3 client/whatsapp/runtime.py
# or
sliderag-whatsapp

4. Start chatting. Supported commands:

• /file <filename> switch the active document
• /status inspect runtime connection/queue/model status

If bridge shows repeated Status: 408 / no QR code:

1. verify WhatsApp Web is reachable from your runtime network,
2. ensure bridge proxy config is effective in the same terminal process,
3. ensure bridge keeps running (do not close after QR).

How to set ALLOW_FROM and STARTUP_NOTIFY_CHAT_ID

After you send one message in QQ/WeChat/WhatsApp, check runtime logs for a line like:

Inbound message: chat_id=..., sender_id=...

Use the sender_id value for your allowlist and startup notification target.

🔗 Related Projects

Project	Description	Link
RAG-Anything	All-in-One RAG Framework	GitHub
nanobot	Ultra-Lightweight Personal AI Assistant	GitHub

🤗 Citation

If you find this project useful, please cite:

@software{sliderag2026,
  title={SlideRAG: PPT-Centric Multimodal RAG for Study Preview and Exam Review},
  author={He Liu,Jiahao Zhang},
  year={2026},
  url={https://github.com/Hitlh/SlideRAG}
}