Skip to main content
Glama

🎬 MCP-Powered Video RAG

用自然语言向你的视频提问——由 Whisper + ChromaDB + Groq + FastMCP 驱动。完全免费,且本地运行。

架构与工作原理

系统架构

Your Video → Whisper (transcription) → ChromaDB (vector store)
                                              ↓
                              Antigravity IDE ← MCP Server ← Your Question
                                              ↓
                                         Groq LLM (free) → Answer + Timestamps

Related MCP server: video-agent-mcp

技术栈(全部免费)

组件

工具

转录

OpenAI Whisper(本地)

嵌入模型

sentence-transformers(本地)

向量数据库

ChromaDB(本地)

MCP 框架

FastMCP

用于问答的 LLM

Groq(免费层)


⚡ 快速开始

1. 安装依赖

uv sync

2. 获取你的免费 Groq API 密钥

前往 https://console.groq.com/ → 创建账户 → 复制 API 密钥。

3. 设置环境

copy .env.example .env
# Edit .env and paste your GROQ_API_KEY

4. 将 MCP 服务器添加到 Antigravity IDE

将 mcp_config.json 的内容复制到你的 Antigravity IDE MCP 设置中。 将 GROQ_API_KEY 更新为你的实际密钥。

5. 将视频放入 videos/ 文件夹

支持格式:.mp4, .mkv, .avi, .mov, .webm, .mp3, .wav


🛠️ 可用的 MCP 工具

工具

描述

ingest_video(video_path)

转录并索引视频文件

search_video(query)

对转录文本进行语义搜索

ask_video(question)

带时间戳的完整 RAG 问答

list_videos()

显示所有已索引的视频

delete_video(video_path)

从索引中移除视频


📖 示例用法(在 Antigravity IDE 中)

ingest_video("videos/my_lecture.mp4")

ask_video("What are the main topics discussed?")

search_video("neural networks explained", n_results=3)

ask_video("What did the speaker say about backpropagation?", video_name="my_lecture.mp4")

⚙️ 配置(.env)

GROQ_API_KEY=your_key_here        # Required — get free at console.groq.com
WHISPER_MODEL=base                # tiny | base | small | medium | large
EMBEDDING_MODEL=all-MiniLM-L6-v2  # local embedding model
GROQ_MODEL=llama-3.1-8b-instant  # Groq model for Q&A
CHROMA_DB_PATH=./chroma_db        # where to persist the vector DB

💡 提示

  • 使用 WHISPER_MODEL=tiny 可获得最快的转录速度(准确度较低)

  • 使用 WHISPER_MODEL=medium 可获得高准确度(速度较慢)

  • 首次运行会下载 Whisper 和嵌入模型(每个约 200MB)——后续运行即时完成

  • ChromaDB 数据在重启后仍会保留在 ./chroma_db/ 中

Available Tools

5 tools
ask_videoA

Ask a natural language question about your videos and get an AI-generated answer.

This tool uses RAG: it retrieves the most relevant transcript chunks from ChromaDB, then sends them as context to a Groq LLM (free tier) to generate a precise answer with timestamps so you can jump directly to the relevant moment in the video.

Inputs: question (required), video_name, n_context_chunks. Returns: result

delete_videoA

Remove a video and all its indexed data from the RAG system.

This deletes all transcript chunks for the specified video from ChromaDB. The original video file is NOT deleted — only the indexed data is removed.

Inputs: video_path (required). Returns: result

ingest_videoA

Ingest a video file into the RAG system.

This transcribes the video using Whisper (locally, for free), splits the transcript into timed chunks, embeds them, and stores them in a local ChromaDB vector database for semantic search.

Inputs: video_path (required). Returns: result

list_videosB

List all videos currently indexed in the RAG system.

Returns a summary of each ingested video including its name, path, and chunk count.

Returns: JSON list of indexed videos with metadata.

Inputs: no named parameters. Returns: result

search_videoB

Semantically search across all indexed video transcripts.

Returns the most relevant timestamped transcript chunks for a given query. You can optionally filter to a specific video by name.

Inputs: query (required), n_results, video_name. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 5 tools

    Disambiguation4/5

    Each tool maps to a distinct action (ingest, search, ask, list, delete). The only potential overlap is search_video vs ask_video, since both retrieve transcript chunks, but the descriptions clearly differentiate raw chunk retrieval from LLM-generated answers with context.

    Naming Consistency5/5

    All five tools follow a strict verb_noun snake_case pattern (ingest_video, search_video, ask_video, list_videos, delete_video). The convention is predictable and uniform.

    Tool Count5/5

    Five tools is well-scoped for a video RAG system. Each tool covers a meaningful, non-redundant part of the workflow with no filler.

    Completeness5/5

    The surface covers the full lifecycle: ingestion, two retrieval modes (search and Q&A), listing indexed content, and deletion. Re-ingesting a video handles the update case, so there are no obvious dead ends.

    Related MCP Connectors

    Related MCP Servers