Add guide for Multi-Agent RAG with Gen2 and Cortex #2750
Conversation
This document provides a comprehensive guide on building a production-grade multi-agent Retrieval-Augmented Generation (RAG) system using Snowflake Gen2 Warehouses and Cortex. It includes architecture, setup instructions, and sample queries.
|
Note: The image uploads happen async, so they may not be uploaded yet if you click this quickly.. try again in a few minutes if so. |
- Complete 6-agent architecture with full SQL/Python implementations - Semantic chunking UDF + Cortex embeddings + external vector DB - Gen2 warehouse optimization with 30-50% performance gains - Production deployment, monitoring, and troubleshooting guides - Based on Medium article about multi-agent RAG with Gen2 warehouses
|
Note: The image uploads happen async, so they may not be uploaded yet if you click this quickly.. try again in a few minutes if so. |
Updated duration formatting throughout the document to include 'Hours'.
|
Note: The image uploads happen async, so they may not be uploaded yet if you click this quickly.. try again in a few minutes if so. |
Updated duration estimates for various sections of the document to reflect more accurate time requirements.
Added overview section for multi-agent RAG system.
|
Note: The image uploads happen async, so they may not be uploaded yet if you click this quickly.. try again in a few minutes if so. |
|
Note: The image uploads happen async, so they may not be uploaded yet if you click this quickly.. try again in a few minutes if so. |
|
Could a maintainer please assign a reviewer for this PR? This guide covers building a production-grade multi-agent RAG system using Snowflake Gen2 Warehouses and Cortex. Thank you! |
sfc-gh-jreini
left a comment
sfc-gh-jreini
left a comment
There was a problem hiding this comment.
Requested some changes
|
|
||
| nltk.download('punkt', quiet=True) | ||
|
|
||
| def chunk_doc(doc_text: str) -> List[str]: |
There was a problem hiding this comment.
Can you use native PARSE_DOCUMENT feature instead of python UDF?
| import numpy as np | ||
| import _snowflake | ||
|
|
||
| def search_vectors(query_embedding: list, top_k: int) -> list: |
There was a problem hiding this comment.
Can you replicate this with cortex search instead of pinecone? Or is the pinecone integration something you're trying to show in the guide?
| ### Create Document Retriever Agent | ||
|
|
||
| ```sql | ||
| CREATE OR REPLACE FUNCTION document_retriever_agent(user_query STRING) |
There was a problem hiding this comment.
This seems more like a retrieval tool than an agent - no agent reasoning is used here, just returning documents. suggest renaming
| ``` | ||
|
|
||
| ## Agent 2: SQL Generator | ||
| Duration: 1.5 Hours |
There was a problem hiding this comment.
very long duration here, intended?
| LANGUAGE SQL | ||
| AS | ||
| $$ | ||
| SELECT SNOWFLAKE.CORTEX.COMPLETE( |
There was a problem hiding this comment.
Use AI_COMPLETE instead of SNOWFLAKE.CORTEX.COMPLETE
| ### Create SQL Generation Agent | ||
|
|
||
| ```sql | ||
| CREATE OR REPLACE FUNCTION sql_generator_agent(user_question STRING) |
There was a problem hiding this comment.
Better to use cortex analyst instead of hand-rolling this sql generator agent. Quality of sql generation will be much higher and more extensible.
| ### Business Logic Layer | ||
|
|
||
| ```sql | ||
| CREATE OR REPLACE FUNCTION semantic_model_agent(entity STRING, operation STRING) |
There was a problem hiding this comment.
What is this agent doing?
| synthesized_response AS ( | ||
| SELECT | ||
| SNOWFLAKE.CORTEX.COMPLETE( | ||
| 'mixtral-8x7b', |
There was a problem hiding this comment.
would recommend using a stronger model here
| AS | ||
| $$ | ||
| WITH agent_results AS ( | ||
| -- Execute all agents in parallel on Gen2 warehouse |
There was a problem hiding this comment.
why execute all agents in parallel if they may not be needed depending on query?
Better to have the coordinator agent do planning/tool selection instead of just using all of the agents at once no matter what.
2 participants
This document provides a comprehensive guide on building a production-grade multi-agent Retrieval-Augmented Generation (RAG) system using Snowflake Gen2 Warehouses and Cortex. It includes architecture, setup instructions, and sample queries.