Issue with AWS Bedrock Knowledge Base Retrieval

Question

We have implemented an AWS Bedrock Knowledge Base solution and are encountering an issue where the response returns:
"Sorry, I'm unable to assist with your request", even though the relevant data is present in the data source and syncing completes successfully.
1.S3 Buckets & Data Upload:
Created two S3 buckets and uploaded required documents.
Automated generation of metadata.json files for each uploaded document with expected attributes.
2.Knowledge Base Setup:
Created a Knowledge Base using:
Default parser
Fixed chunking strategy
Embedding model: amazon.titan-embed-text-v2 from Bedrock.
Linked both buckets as data sources and triggered the sync process.
3.Knowledge Base Querying:
Accessing the knowledge base using the assigned knowledgeBaseId.
Retrieval is configured with HYBRID search and numberOfResults set to 10.
Metadata filtering has been added to enhance result accuracy.
Base Configuration Snippet:

base_config = {
  "type": "KNOWLEDGE_BASE",
  "knowledgeBaseConfiguration": {
    "knowledgeBaseId": merged_config['kb_md_id'],
    "modelArn": arn,
    "retrievalConfiguration": {
      "vectorSearchConfiguration": {
        "numberOfResults": 10,
        "overrideSearchType": 'HYBRID'
      }
    },
    "generationConfiguration": {
      "promptTemplate": {
        "textPromptTemplate": promptTemplate
      },
      "inferenceConfig": {
        "textInferenceConfig": {
          "temperature": 0.3,
          "topP": 1,
          "maxTokens": 2048,
          "stopSequences": ["observation"]
        }
      }
    }
  }
}
filter_conditions = []

if matched_brands:
  filter_conditions.append({
    "equals": {
      "key": "brandname",
      "value": matched_brands[0]
    }
  })

if part_numbers:
  part_number_conditions = [
    {
      "stringContains": {
        "key": "parts",
        "value": f'"partnumber": "{p}"'
      }
    } for p in part_numbers
  ]
  filter_conditions.extend(part_number_conditions)

if len(filter_conditions) > 1:
  base_config["knowledgeBaseConfiguration"]["retrievalConfiguration"]["vectorSearchConfiguration"]["filter"] = {
    "andAll": filter_conditions
  }
elif filter_conditions:
  base_config["knowledgeBaseConfiguration"]["retrievalConfiguration"]["vectorSearchConfiguration"]["filter"] = \
    filter_conditions[0]

Note: Validated the metadata structure for filter enablement

re:Post Agent · Answer

Based on the information you've provided, it seems you're experiencing an issue with your AWS Bedrock Knowledge Base where queries are not returning expected results, despite having relevant data in your data source. Here are some potential reasons and solutions for this problem:

1. Metadata Filtering: Ensure that your metadata is correctly formatted and matches the structure expected by Bedrock. For numerical values, avoid using quotes. For example, use `"caseId": 2` instead of `"caseId": "2"`. This can affect how Bedrock processes the metadata for filtering.

2. Retrieval Configuration: Your retrieval configuration might need adjustment. Make sure the `numberOfResults` parameter is set high enough to capture all relevant chunks. You might want to increase this value to ensure you're not missing any relevant information.

3. Query Format: The structure of your retrieval query is crucial. Make sure it follows the correct format as specified in the Bedrock documentation. For example:

```json
{
  "retrievalQuery": {
    "text": "Your query here"
  },
  "retrievalConfiguration": {
    "vectorSearchConfiguration": {
      "numberOfResults": 100
    }
  }
}
```

4. Chunking Strategy: Review your fixed chunking strategy. If chunks are too small or too large, it might affect the relevance of retrieved information.

5. Embedding Model Access: Verify that you have proper access to the Titan embedding model (amazon.titan-embed-text-v2). Sometimes, even with general Bedrock access, you might need to explicitly request access to specific models.

6. Sync Process: Ensure that your data sync process completed successfully. Check for any sync errors in the Bedrock console or CloudWatch logs.

7. IAM Permissions: Confirm that your IAM roles have the necessary permissions to access all required services, including S3, Bedrock, and any other integrated services.

8. Region Compatibility: Make sure your S3 buckets and Bedrock resources are in the same AWS region.

If you've verified all these points and are still experiencing issues, you may want to try the following:

- Use the "Test knowledge base" feature in the Bedrock console to isolate whether the problem is with your application code or the knowledge base itself.
- Review CloudWatch logs for any error messages or unexpected behavior.
- Consider creating a new knowledge base with a subset of your data to test if the issue persists.

If none of these solutions resolve your issue, it may be worth contacting AWS Support for more detailed troubleshooting specific to your account and configuration.  
**Sources**  
[Issue with metadata filtering in Amazon Bedrock | AWS re:Post](https://repost.aws/questions/QU3eHU3zRATX6sYPuWVEHGsw/issue-with-metadata-filtering-in-amazon-bedrock)  
[Missing records form Bedrock Knowledge Base Query , maybe chunking issue | AWS re:Post](https://repost.aws/questions/QU7MF_J9H6Sm-QC45ubWPzzg/missing-records-form-bedrock-knowledge-base-query-maybe-chunking-issue)  
[Data sync failed error in Amazon Bedrock Knowledgebase | AWS re:Post](https://repost.aws/questions/QUhiVO9lYTRRaAgo2MEmA8TA/data-sync-failed-error-in-amazon-bedrock-knowledgebase)  
[Bedrock knowledge base data source sync is not working | AWS re:Post](https://repost.aws/questions/QUjagkPYWnTu299Hr7JMnHIQ/bedrock-knowledge-base-data-source-sync-is-not-working)

Issue with AWS Bedrock Knowledge Base Retrieval

Add your answer

Relevant content