Information Bases for Amazon Bedrock now helps metadata filtering to enhance retrieval accuracy

At AWS re:Invent 2023, we introduced the overall availability of Information Bases for Amazon Bedrock. With Information Bases for Amazon Bedrock, you’ll be able to securely join basis fashions (FMs) in Amazon Bedrock to your organization knowledge utilizing a completely managed Retrieval Augmented Technology (RAG) mannequin.

For RAG-based functions, the accuracy of the generated responses from FMs rely upon the context supplied to the mannequin. Contexts are retrieved from vector shops based mostly on consumer queries. Within the lately launched function for Information Bases for Amazon Bedrock, hybrid search, you’ll be able to mix semantic search with key phrase search. Nevertheless, in lots of conditions, you might have to retrieve paperwork created in an outlined interval or tagged with sure classes. To refine the search outcomes, you’ll be able to filter based mostly on doc metadata to enhance retrieval accuracy, which in flip results in extra related FM generations aligned along with your pursuits.

On this put up, we focus on the brand new customized metadata filtering function in Information Bases for Amazon Bedrock, which you should utilize to enhance search outcomes by pre-filtering your retrievals from vector shops.

Metadata filtering overview

Previous to the discharge of metadata filtering, all semantically related chunks as much as the pre-set most could be returned as context for the FM to make use of to generate a response. Now, with metadata filters, you’ll be able to retrieve not solely semantically related chunks however a well-defined subset of these related chucks based mostly on utilized metadata filters and related values.

With this function, now you can provide a customized metadata file (every as much as 10 KB) for every doc within the information base. You’ll be able to apply filters to your retrievals, instructing the vector retailer to pre-filter based mostly on doc metadata after which seek for related paperwork. This manner, you’ve management over the retrieved paperwork, particularly in case your queries are ambiguous. For instance, you should utilize authorized paperwork with comparable phrases for various contexts, or motion pictures which have an identical plot launched in several years. As well as, by decreasing the variety of chunks which might be being searched over, you obtain efficiency benefits like a discount in CPU cycles and price of querying the vector retailer, along with enchancment in accuracy.

To make use of the metadata filtering function, you want to present metadata recordsdata alongside the supply knowledge recordsdata with the identical identify because the supply knowledge file and .metadata.json suffix. Metadata may be string, quantity, or Boolean. The next is an instance of the metadata file content material:

{
    "metadataAttributes" : { 
        "tag" : "challenge EVE",
        "yr" :  2016,
        "crew": "ninjas"
    }
}

The metadata filtering function of Information Bases for Amazon Bedrock is obtainable in AWS Areas US East (N. Virginia) and US West (Oregon).

The next are widespread use instances for metadata filtering:

• Doc chatbot for a software program firm – This enables customers to search out product data and troubleshooting guides. Filters on the working system or utility model, for instance, might help keep away from retrieving out of date or irrelevant paperwork.
• Conversational search of a company’s utility – This enables customers to go looking via paperwork, kanbans, assembly recording transcripts, and different property. Utilizing metadata filters on work teams, enterprise models, or challenge IDs, you’ll be able to personalize the chat expertise and enhance collaboration. An instance could be, “What’s the standing of challenge Sphinx and dangers raised,” the place customers can filter paperwork for a selected challenge or supply kind (akin to e-mail or assembly paperwork).
• Clever seek for software program builders – This enables builders to search for data of a selected launch. Filters on the discharge model, doc kind (akin to code, API reference, or subject) might help pinpoint related paperwork.

Answer overview

Within the following sections, we exhibit tips on how to put together a dataset to make use of as a information base, after which question with metadata filtering. You’ll be able to question utilizing both the AWS Administration Console or SDK.

Put together a dataset for Information Bases for Amazon Bedrock

For this put up, we use a pattern dataset about fictional video video games for instance tips on how to ingest and retrieve metadata utilizing Information Bases for Amazon Bedrock. If you wish to observe alongside in your individual AWS account, obtain the file.

If you wish to add metadata to your paperwork in an current information base, create the metadata recordsdata with the anticipated filename and schema, then skip to the step to sync your knowledge with the information base to begin the incremental ingestion.

In our pattern dataset, every recreation’s doc is a separate CSV file (for instance, s3://$bucket_name/video_game/$game_id.csv) with the next columns:

title, description, genres, yr, writer, rating

Every recreation’s metadata has the suffix .metadata.json (for instance, s3://$bucket_name/video_game/$game_id.csv.metadata.json) with the next schema:

{
  "metadataAttributes": {
    "id": quantity, 
    "genres": string,
    "yr": quantity,
    "writer": string,
    "rating": quantity
  }
}

Create a information base for Amazon Bedrock

For directions to create a brand new information base, see Create a information base. For this instance, we use the next settings:

• On the Arrange knowledge supply web page, beneath Chunking technique, choose No chunking, since you’ve already preprocessed the paperwork within the earlier step.
• Within the Embeddings mannequin part, select Titan G1 Embeddings – Textual content.
• Within the Vector database part, select Fast create a brand new vector retailer. The metadata filtering function is obtainable for all supported vector shops.

Synchronize the dataset with the information base

After you create the information base, and your knowledge recordsdata and metadata recordsdata are in an Amazon Easy Storage Service (Amazon S3) bucket, you can begin the incremental ingestion. For directions, see Sync to ingest your knowledge sources into the information base.

Question with metadata filtering on the Amazon Bedrock console

To make use of the metadata filtering choices on the Amazon Bedrock console, full the next steps:

1. On the Amazon Bedrock console, select Information bases within the navigation pane.
2. Select the information base you created.
3. Select Check information base.
4. Select the Configurations icon, then broaden Filters.
5. Enter a situation utilizing the format: key = worth (for instance, genres = Technique) and press Enter.
6. To alter the important thing, worth, or operator, select the situation.
   
7. Proceed with the remaining situations (for instance, (genres = Technique AND yr >= 2023) OR (ranking >= 9))
   
8. When completed, enter your question within the message field, then select Run.

For this put up, we enter the question “A technique recreation with cool graphic launched after 2023.”

Question with metadata filtering utilizing the SDK

To make use of the SDK, first create the consumer for the Brokers for Amazon Bedrock runtime:

import boto3

bedrock_agent_runtime = boto3.consumer(
    service_name = "bedrock-agent-runtime"
)

Then assemble the filter (the next are some examples):

# genres = Technique
single_filter= {
    "equals": {
        "key": "genres",
        "worth": "Technique"
    }
}

# genres = Technique AND yr >= 2023
one_group_filter= {
    "andAll": [
        {
            "equals": {
                "key": "genres",
                "value": "Strategy"
            }
        },
        {
            "GreaterThanOrEquals": {
                "key": "year",
                "value": 2023
            }
        }
    ]
}

# (genres = Technique AND yr >=2023) OR rating >= 9
two_group_filter = {
    "orAll": [
        {
            "andAll": [
                {
                    "equals": {
                        "key": "genres",
                        "value": "Strategy"
                    }
                },
                {
                    "GreaterThanOrEquals": {
                        "key": "year",
                        "value": 2023
                    }
                }
            ]
        },
        {
            "GreaterThanOrEquals": {
                "key": "rating",
                "worth": "9"
            }
        }
    ]
}

Go the filter to retrievalConfiguration of the Retrieval API or RetrieveAndGenerate API:

retrievalConfiguration={
        "vectorSearchConfiguration": {
            "filter": metadata_filter
        }
    }

The next desk lists just a few responses with completely different metadata filtering situations.

Along with customized metadata, you may also filter utilizing S3 prefixes (which is a built-in metadata, so that you don’t want to offer any metadata recordsdata). For instance, in the event you set up the sport paperwork into prefixes by writer (for instance, s3://$bucket_name/video_game/$writer/$game_id.csv), you’ll be able to filter with the precise writer (for instance, neo_tokyo_games) utilizing the next syntax:

publisher_filter = {
    "startsWith": {
                    "key": "x-amz-bedrock-kb-source-uri",
                    "worth": "s3://$bucket_name/video_game/neo_tokyo_games/"
                }
}

Clear up

To scrub up your assets, full the next steps:

1. Delete the information base:
   1. On the Amazon Bedrock console, select Information bases beneath Orchestration within the navigation pane.
   2. Select the information base you created.
   3. Be aware of the AWS Id and Entry Administration (IAM) service function identify within the Information base overview part.
   4. Within the Vector database part, pay attention to the gathering ARN.
   5. Select Delete, then enter delete to substantiate.
2. Delete the vector database:
   1. On the Amazon OpenSearch Service console, select Collections beneath Serverless within the navigation pane.
   2. Enter the gathering ARN you saved within the search bar.
   3. Choose the gathering and selected Delete.
   4. Enter affirm within the affirmation immediate, then select Delete.
3. Delete the IAM service function:
   1. On the IAM console, select Roles within the navigation pane.
   2. Seek for the function identify you famous earlier.
   3. Choose the function and select Delete.
   4. Enter the function identify within the affirmation immediate and delete the function.
4. Delete the pattern dataset:
   1. On the Amazon S3 console, navigate to the S3 bucket you used.
   2. Choose the prefix and recordsdata, then select Delete.
   3. Enter completely delete within the affirmation immediate to delete.

Conclusion

On this put up, we lined the metadata filtering function in Information Bases for Amazon Bedrock. You discovered tips on how to add customized metadata to paperwork and use them as filters whereas retrieving and querying the paperwork utilizing the Amazon Bedrock console and the SDK. This helps enhance context accuracy, making question responses much more related whereas reaching a discount in price of querying the vector database.

For added assets, check with the next:

In regards to the Authors

Corvus Lee is a Senior GenAI Labs Options Architect based mostly in London. He’s obsessed with designing and creating prototypes that use generative AI to unravel buyer issues. He additionally retains up with the newest developments in generative AI and retrieval strategies by making use of them to real-world eventualities.

Ahmed Ewis is a Senior Options Architect at AWS GenAI Labs, serving to prospects construct generative AI prototypes to unravel enterprise issues. When not collaborating with prospects, he enjoys taking part in together with his youngsters and cooking.

Chris Pecora is a Generative AI Knowledge Scientist at Amazon Net Providers. He’s obsessed with constructing modern merchandise and options whereas additionally specializing in customer-obsessed science. When not working experiments and maintaining with the newest developments in GenAI, he loves spending time together with his youngsters.