skygazer42/GustoBot

What it solves

GustoBot is a specialized AI customer service system for Chinese recipes. It addresses the difficulty of exploring deep recipe knowledge—such as regional characteristics, nutritional principles, and historical anecdotes—which traditional keyword searches cannot handle due to the mixture of highly structured data (ingredients, steps) and unstructured cultural descriptions.

How it works

The system uses a three-layer Multi-Agent architecture built with LangGraph:

1. Main Routing Layer (L1): Uses heuristic keywords and LLM-based intent recognition to route queries to the appropriate module.
2. Multi-Tool Subgraph Layer (L2): Decomposes complex tasks using a Map-Reduce pattern. It employs a Planner node to break down questions into sub-tasks, which are then executed in parallel by specific tools.
3. Atomic Tool Layer (L3): Executes the actual data retrieval using a diverse stack including Neo4j (knowledge graphs), MySQL (statistical analysis via Text2SQL), Milvus and PostgreSQL pgvector (semantic and structured vector search), and external search APIs.

It features a "PostgreSQL-first" fallback strategy (Structured $\rightarrow$ Vector $\rightarrow$ External) and a Reranker with dual thresholds to ensure high answer precision.

Who it’s for

• Culinary enthusiasts seeking detailed recipe and cultural information.
• Developers looking for a professional, transferable template for building vertical-domain AI agents (e.g., for encyclopedias, legal or medical consultation).

Highlights

• Multi-Agent Orchestration: Uses LangGraph for complex task decomposition and parallel tool execution.
• Hybrid RAG: Combines GraphRAG (Neo4j), Text2SQL (MySQL), and vector search (Milvus/PostgreSQL).
• L1-L3 Architecture: A clear separation of concerns that makes the system easily adaptable to other domains.
• Multimodal Capabilities: Supports image recognition for dishes and image generation via CogView-4.
• Safety Guardrails: Integrated range checks to prevent out-of-scope queries.
• Source Traceability: Every answer is annotated with its knowledge source.