Finance ops · Production

From RAG to fabric: Lessons learned from building real-world RAGs at GenAIIC – Part 2

The problem

Building RAG systems that handle heterogeneous data formats—structured tables, unstructured text, and images—requires distinct retrieval and processing strategies per data type, which a single uniform approach cannot address, especially since LLMs perform poorly on raw tabular data.

Workflow diagram · grounded in source

User query submitted

trigger

“a router is a component that directs incoming user queries to the appropriate processing pipeline based on the query's nature and the required data type”

Router intent detection

routing

“The router accomplishes this through intent detection, analyzing the query to determine the type of data and analysis required to answer it”

LLM code generation for structured data

ai_action

“asking the LLM to write Python (if the data is stored in a CSV, Excel, or Parquet file) or SQL (if the data is stored in a SQL database) code that performs the required analysis”

Multimodal embedding retrieval

ai_action

“a multimodal embedding model such as Amazon Titan Multimodal Embeddings, which can embed both images and text into a shared vector space”

Coherent response delivered

output

“create intelligent systems that can understand queries, retrieve relevant information from structured and unstructured data sources, and provide coherent responses”

Reported outcome

By employing intent detection routers, LLM code generation, and multimodal embeddings, the GenAIIC team built intelligent RAG systems spanning oil and gas, financial, industrial, and ecommerce use cases that deliver coherent responses across heterogeneous data sources.

Reported metrics

Claude Sonnet 3.5 HumanEval code-generation accuracy92%

Reported stack

Amazon BedrockClaude HaikuClaude Sonnet 3.5Claude Sonnet 3Amazon Titan Multimodal EmbeddingsAmazon Titan Embedding Text v2OpenSearchLlama IndexLangChainBedrock Converse API

Source

https://aws.amazon.com/blogs/machine-learning/from-rag-to-fabric-lessons-learned-from-building-real-world-rags-at-genaiic-part-2?tag=soumet-20

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

By employing intent detection routers, LLM code generation, and multimodal embeddings, the GenAIIC team built intelligent RAG systems spanning oil and gas, financial, industrial, and ecommerce use cases that deliver c…

What tools did this team use?

Amazon Bedrock, Claude Haiku, Claude Sonnet 3.5, Claude Sonnet 3, Amazon Titan Multimodal Embeddings, Amazon Titan Embedding Text v2, OpenSearch, Llama Index, LangChain, Bedrock Converse API.

What results were reported?

Claude Sonnet 3.5 HumanEval code-generation accuracy: 92% (source-reported, not independently verified).

How is this finance ops AI workflow structured?

User query submitted → Router intent detection → LLM code generation for structured data → Multimodal embedding retrieval → Coherent response delivered.