Amazon Bedrock knowledge base now supports metadata filtering to improve search accuracy

At AWS re:Invent 2023, we introduced the final availability of the Amazon Bedrock information base. Via the Amazon Bedrock Data Base, you may securely join base fashions (FM) in Amazon Bedrock to your organization knowledge utilizing absolutely managed Retrieval Augmented Era (RAG) fashions.

For RAG-based purposes, the accuracy of the response produced by FM relies on the context offered to the mannequin. Retrieve context from vector storage based mostly on consumer question. Within the not too long ago launched hybrid search characteristic of Amazon Bedrock’s information base, you may mix semantic search with key phrase search. Nevertheless, in lots of circumstances you could must retrieve recordsdata that have been created inside an outlined time interval or tagged with sure classes. To refine your search outcomes, you may filter based mostly on file metadata to enhance search accuracy and produce related FMs that higher fit your pursuits.

On this article, we talk about the brand new customized metadata filtering characteristic within the Amazon Bedrock Data Base, which you need to use to enhance search outcomes by pre-filtering searches within the vector retailer.

Metadata filtering overview

Earlier than metadata filtering is revealed, all semantically related chunks (as much as a preset most) are handed again as context for FM to make use of to generate responses. Now, with metafilters, you cannot solely retrieve semantically associated chunks, but in addition a well-defined subset of these related chunks based mostly on the utilized metafilter and related worth.

With this characteristic, now you can present a customized metadata file (as much as 10 KB per metadata file) for every doc in your information base. You’ll be able to apply a filter to your search, instructing the vector retailer to pre-filter based mostly on doc metadata after which seek for related paperwork. This fashion you may management which paperwork are retrieved, particularly when your question is ambiguous. For instance, you would use authorized paperwork with related terminology in numerous contexts, or films with related plots launched in numerous years. Moreover, by decreasing the variety of blocks being searched, along with enhancing accuracy, you achieve efficiency advantages corresponding to decreased CPU cycles and question vector storage prices.

To make use of metadata filtering, you could present a metadata file with the identical title because the supply knowledge file subsequent to the supply knowledge file, and .metadata.json suffix. Metadata may be strings, numbers, or Boolean values. The next is an instance of metadata file content material:

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

Metadata filtering for the Amazon Bedrock Data Base is obtainable within the AWS Areas US East (N. Virginia) and US West (Oregon).

The next are widespread use circumstances for metadata filtering:

• A software program firm’s doc chatbot – This enables customers to seek out product data and troubleshooting guides. For instance, working system or utility model filters may also help keep away from retrieving outdated or irrelevant paperwork.
• Conversational seek for organizational apps – This enables customers to go looking paperwork, dashboards, assembly minutes and different property. You’ll be able to personalize the chat expertise and enhance collaboration utilizing metadata filters on workgroups, enterprise models, or challenge IDs. For instance, “What’s the standing and threat of Sphinx tasks,” the place customers can filter recordsdata for particular tasks or supply varieties (corresponding to emails or assembly paperwork).
• Good search software program developer – This enables builders to seek out details about a selected model. Filters on releases, doc varieties (corresponding to code, API reference, or points) may also help discover related paperwork.

Answer overview

Within the following sections, we exhibit how one can put together a dataset to be used as a information base after which question it utilizing metadata filtering. You’ll be able to question utilizing the AWS Administration Console or SDKs.

Put together an information set for the Amazon Bedrock information base

On this article, we use a pattern dataset a few fictional online game as an example how one can use the Amazon Bedrock information base to extract and retrieve metadata. If you wish to use your personal AWS account, obtain the file.

If you’re including metadata to a doc in an present information base, create a metadata file with the anticipated file title and schema, then skip to the steps to synchronize the fabric with the information base to start incremental ingestion.

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

title, description, genres, yr, writer, rating

Every sport’s metadata has a suffix .metadata.json (For instance, s3://$bucket_name/video_game/$game_id.csv.metadata.json) has the next structure:

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

Construct a information base for Amazon Bedrock

For directions on organising a brand new information base, see Constructing a Data Base. For this instance, we use the next settings:

• superior Set knowledge supply web page, subsequent chunking techniqueselect No chunkingsince you already preprocessed the file within the earlier step.
• inside embedded mannequin half, choose Titan G1 Embed – Textual content.
• inside vector database half, choose Rapidly create new vector shops. Metadata filtering works on all supported vector shops.

Synchronize datasets with information base

As soon as the information base is created and your knowledge recordsdata and metadata recordsdata are in an Amazon Easy Storage Service (Amazon S3) bucket, you can begin incremental retrieval. For directions, see Sync to convey sources into the information base.

Question utilizing 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 base Within the navigation pane.
2. Choose the information base you created.
3. select Take a look at information base.
4. select Configuration icon and increase filter.
5. Enter the factors utilizing the next format: key = worth (for instance, style = technique) and press Enter.
6. To vary keys, values, or operators, choose a situation.
   
7. Proceed with the remainder of the factors (e.g. (style = technique AND yr >= 2023) OR (score >= 9))
   
8. When completed, enter your question within the message field and choose working.

For this text, we entered the question “Technique video games with cool graphics launched after 2023”.

Use SDK to carry out metadata filtering queries

To make use of the SDK, first arrange the shopper for the Brokers for Amazon Bedrock execution part:

import boto3

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

Then assemble the filter (listed below 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"
            }
        }
    ]
}

Cross the filter to retrievalConfiguration Retrieval API or RetrieveAndGenerate API:

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

The desk beneath lists some responses with totally different metadata filters.

Along with customized metadata, you can too filter utilizing S3 prefixes (that is built-in metadata, so that you needn’t present any metadata recordsdata). For instance, when you set up your sport recordsdata by writer into prefixes (e.g. s3://$bucket_name/video_game/$writer/$game_id.csv), you may filter by particular publishers (for instance, neo_tokyo_games) makes use of 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 sources, full the next steps:

1. Delete a information base:
   1. On the Amazon Bedrock console, select information base beneath Organize Within the navigation pane.
   2. Choose the information base you created.
   3. Make an observation of the AWS Id and Entry Administration (IAM) service function title Data Base Overview half.
   4. inside vector database part, write down the gathering ARN.
   5. select deletethen enter Delete to verify.
2. Delete vector library:
   1. On the Amazon OpenSearch Service console, select accumulate beneath No server Within the navigation pane.
   2. Enter the gathering ARN you saved within the search bar.
   3. choose set and choose delete.
   4. Enter verify within the affirmation immediate and choose delete.
3. Delete the IAM service function:
   1. On the IAM console, select Position Within the navigation pane.
   2. Seek for the character title you famous earlier.
   3. Select a job and select delete.
   4. Enter the function title within the affirmation immediate and delete the function.
4. Delete a pattern knowledge set:
   1. On the Amazon S3 console, navigate to the S3 bucket you’re utilizing.
   2. Choose the prefix and file, then choose delete.
   3. Enter everlasting delete within the affirmation immediate to delete.

in conclusion

On this article, we introduce the metadata filtering capabilities within the Amazon Bedrock information base. You learn to add customized metadata to recordsdata and use it as a filter, whereas retrieving and querying recordsdata utilizing the Amazon Bedrock console and SDKs. This helps enhance contextual accuracy, making question responses extra related whereas decreasing the price of querying the vector repository.

For extra sources, see the next sources:

Concerning the creator

Corvus plum is a Senior Options Architect at GenAI Labs based mostly in London. He’s keen about designing and growing prototypes that use generative synthetic intelligence to unravel buyer issues. He additionally retains up with the most recent developments in generative synthetic intelligence and retrieval strategies by making use of them to real-world eventualities.

Ahmed Yuis He’s a senior options architect on the AWS GenAI Lab, helping clients in constructing generative AI prototypes to unravel enterprise issues. When not working with clients, he enjoys taking part in along with his youngsters and cooking.

Chris Pecora is a Generative Synthetic Intelligence Knowledge Scientist at Amazon Internet Companies. He’s keen about constructing modern merchandise and options whereas additionally specializing in customer-centric science. When not conducting experiments and studying in regards to the newest developments in GenAI, he enjoys spending time along with his youngsters.