cinematch-ai / README.md
dbadeev's picture
Update README.md
926c371 verified
|
Raw
History Blame Contribute Delete
4.1 kB
---
title: CineMatch AI (Powered by Modal & Nebius)
emoji: 🎬
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 6.0.1
app_file: app.py
pinned: false
license: mit
short_description: Multi-Agent RAG for Semantic Movie Discovery
tags:
- mcp-in-action-track-01 # Уточни номер трека в правилах, обычно это 01 или текущий месяц
- agents
- rag
- modal
- nebius
- search
---
# 🎬 CineMatch AI: Semantic Movie Discovery via Multi-Agent RAG
### 🏆 Submission for MCP 1st Birthday Hackathon
**Track:** MCP in Action
**Special Category:** Modal Innovation Award
---
### 🚀 Elevator Pitch
**Describe your story, find its cinematic soulmates.**
CineMatch AI is a creative partner for movie discovery. Instead of keywords, users describe their own original story ideas, dreams, or real-life situations (50+ words). A team of **5 Autonomous Agents** collaborates to refine this narrative, retrieve semantic matches from a vector database (FAISS running on **Modal GPU**), and expertly justify *why* these movies fit the user's concept.
Crucially, the system features a **Shadow Evaluation Agent (AgentOps)** that runs silently in the background via MCP, auditing recommendations for hallucinations and logical coherence.
---
### 🛠️ Architecture & MCP Integration
The system uses a **Hierarchical Multi-Agent Architecture** orchestrated on **Modal (Serverless GPU)** and powered by **Llama-3.3-70B** (via Nebius AI).
#### The Agentic Pipeline:
1. **👮‍♂️ Coordinator Agent:** Validates input and manages the conversation state machine. Uses **Context Engineering** to guide users.
2. **📝 Editor & 🎞️ Critic Agents:** Refine the user's raw text into a professional synopsis optimized for embedding.
3. **🔍 Retriever Agent (MCP Tool User):**
* Acts as a bridge to our Vector Database.
* Calls the semantic search tool (hosted on Modal) which queries a **FAISS index** of thousands of movies.
* Uses **all-MiniLM-L6-v2** for embedding generation.
4. **🏆 Expert Agent:** Performs parallel evaluation of candidates and generates persuasive justifications.
5. **🕵️ Shadow Evaluator (AgentOps):**
* Demonstrates the power of MCP for monitoring.
* Runs asynchronously to audit the expert's output against ground truth metadata, detecting hallucinations without disrupting the UX.
---
### 🔌 Dual-Layer Architecture (MCP Integration)
CineMatch AI features a unique dual-layer approach to balance speed and standardization:
1. **Direct RPC Layer:** Used by the Gradio Web UI for **low-latency** user interactions. It communicates directly with Modal functions via optimized RPC calls.
2. **MCP Layer (`tools/`):** A fully compliant **Model Context Protocol** implementation. This allows external agents (like Claude Desktop or other MCP clients) to discover and connect to our Movie Search Engine as a standard tool, fulfilling the hackathon's interoperability requirements.
---
### 🏗️ Technical Stack
* **Infrastructure:** [Modal](https://modal.com) (Serverless Python, GPU A10G for FAISS/Embeddings).
* **LLM:** Meta Llama 3.3 70B Instruct (via Nebius AI).
* **Protocol:** Agents communicate using structured tool calls (MCP pattern).
* **Interface:** Gradio 6.
### 🎥 Demo Video
[[LINK TO YOUR VIDEO HERE](https://drive.google.com/file/d/16s1PH7T3sfaGebbhyHfBzCOG0cFF2cJA/view?usp=sharing)]
### 🔗 Social Media Post
[[LINK TO YOUR SOCIAL MEDIA POST HERE](https://www.linkedin.com/posts/dmitry-badeev-7234a91a9_ai-multiagent-rag-activity-7401030457754091520-gpLf?utm_source=share&utm_medium=member_desktop&rcm=ACoAADCpKssB9RwEKiQBmAkxhLZgt_56k2Ikd-M)]
---
### 🚀 How to Use
1. Enter a detailed description of a story, dream, or life situation (e.g., *"A story about a lonely botanist on Mars who learns to communicate with rocks. He was ..." (50+ words)*).
2. Watch the agents collaborate in real-time.
3. Receive 3 highly relevant movie recommendations with expert justifications.
4. (Optional) Check the logs to see the Shadow Evaluator's quality audit.