AzureOpen AI With Multiple CSV

Question

Hi Everyone,

I’m currently working on a setup using Azure Blob Storage where I upload multiple CSV files. These files are being indexed through Azure Cognitive Search (Standard tier, not Basic).

After indexing, I connect the search index as a data source to an Azure AI / GPT model, but the results are not accurate and don’t return the expected information.

I would really appreciate any guidance on the best practices for:

Structuring CSV files for better indexing

Configuring Cognitive Search indexers and skillsets

Improving search quality or vector retrieval

Ensuring accurate responses when the index is connected to Azure OpenAI / AI Studio

If anyone has experience or recommendations for improving accuracy in this setup, please let me know.

Thank you!

Answer 1

Hello Ali Rizvi,

Welcome to the Microsoft Q&A and thank you for posting your questions here.

I understand that you would like to work efficiently using AzureOpen AI With Multiple CSV.

With the steps below you will be able achieve your aim:

1. Put into consideration your architecture: Pipeline: Blob Storage > Cognitive Search (Hybrid Index) > Azure OpenAI (RTR Pattern)
2. Structuring CSV Files is very important:
   • Ensure consistent schema across all CSVs.
   • Use descriptive headers and normalize data types.
   • Add a unique id column for each row.
   Use this link - https://learn.microsoft.com/en-us/azure/search/search-howto-indexing-azure-blob-storage for more details.
3. Your Cognitive Search Index Design should be: Create fields:

      `content` (searchable, retrievable)
   
      `embedding` (vector field for embeddings)
   

   Enable Semantic Search and Hybrid Search. Check this on https://learn.microsoft.com/en-us/azure/search/semantic-search-overview
4. Embedding Generation:
   • Use Azure OpenAI text-embedding-ada-002 or text-embedding-3-large.
   • Chunk text into 500–1000 tokens before embedding.
   • Store embeddings in Cognitive Search vector fields. For more details - https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/embeddings
5. On RTR Pattern Implementation navigate:
   • Query Cognitive Search > Retrieve top N results > Pass as context to GPT. Use system prompt: “Answer based only on the provided context. If unsure, say ‘Not found.’” Keep temperature=0 for factual accuracy. use this link https://learn.microsoft.com/en-us/azure/ai-services/openai/use-your-data-overview for more details.
6. For the skillsets:
   • Add Text Analytics for entity/key phrase extraction.
   • Use Custom Skills for domain-specific enrichment.
   https://learn.microsoft.com/en-us/azure/search/cognitive-search-skillset>

Finally, you need to put into consideration:

• Cost by optimize embedding generation by batching.
• Latency by using cache frequent queries.
• Context overflow by limit to top 3–5 chunks.

I hope this is helpful! Do not hesitate to let me know if you have any other questions or clarifications.

---

Please don't forget to close up the thread here by upvoting and accept it as an answer if it is helpful.