Amazon Bedrock’s information base now helps metadata filtering to enhance retrieval accuracy

in AWS re:Invent In 2023, we introduced the final availability of the Amazon Bedrock information base. The Amazon Bedrock Information Base permits you to securely join your Amazon Bedrock Basis Mannequin (FM) to your enterprise information utilizing a completely managed Acquisition, Augmentation and Era (RAG) mannequin.

For RAG-based purposes, the accuracy of the responses generated from FM relies on the context supplied to the mannequin. Context is retrieved from the vector retailer based mostly on the person’s question. Hybrid search, a just lately launched information base function in Amazon Bedrock, permits you to mix semantic and key phrase searches. Nonetheless, in lots of conditions, you might wish to retrieve paperwork created throughout an outlined time interval or tagged with a sure class. To slender your search outcomes, you may refine your search by filtering based mostly on doc metadata. This ends in a extra related FM era tailor-made to your pursuits.

This publish describes the brand new customized metadata filtering function in Amazon Bedrock’s information base. This function permits you to enhance your search outcomes by pre-filtering retrieval from the vector retailer.

Overview of metadata filtering

Previous to the discharge of metadata filtering, all semantically associated chunks as much as a preset most have been returned as context for the FM used to generate the response. Metadata filters can help you retrieve not solely semantically associated chunks, but additionally a well-defined subset of these associated chunks based mostly on the utilized metadata filter and related values. .

This function permits you to present customized metadata information (as much as 10 KB every) for every doc in your information base. You possibly can apply filters to your search to inform the vector retailer to pre-filter and discover related paperwork based mostly on doc metadata. This offers you management over the paperwork retrieved, particularly in case your question is ambiguous. For instance, you should utilize authorized paperwork with related phrases in several contexts, or films with related plots launched in several years. Moreover, by decreasing the variety of chunks searched, along with improved accuracy, there are additionally efficiency advantages reminiscent of diminished CPU cycles and vector retailer question prices.

To make use of the metadata filtering function, you have to present a metadata file with the identical title because the supply information file together with the supply information file. .metadata.json suffix. Metadata might be strings, numbers, or Boolean values. Under is an instance of the contents of a metadata file.

{
    "metadataAttributes" : { 
        "tag" : "challenge EVE",
        "12 months" :  2016,
        "staff": "ninjas"
    }
}

Amazon Bedrock’s information base metadata filtering function is obtainable within the US East (N. Virginia) and US West (Oregon) AWS Areas.

Frequent use instances for metadata filtering embody:

• Doc chatbot for software program corporations – This permits customers to seek out product data and troubleshooting guides. For instance, filters based mostly on working system or utility model can assist you keep away from retrieving outdated or irrelevant paperwork.
• Conversational seek for your group’s purposes – This permits customers to seek for paperwork, playing cards, assembly notes, and different belongings. You possibly can personalize the chat expertise and enhance collaboration by utilizing metadata filters for work teams, enterprise items, or challenge IDs. Examples embody “What’s the standing of my challenge Sphinx and what dangers does it pose?” which permits customers to filter paperwork for particular initiatives or supply varieties (reminiscent of emails or assembly supplies).
• Clever seek for software program builders – This permits builders to search for data for a particular launch. Filters by launch model, doc sort (code, API reference, concern, and so forth.) assist pinpoint related documentation.

Resolution overview

The following part exhibits you the right way to put together a dataset to be used as a information base and run queries utilizing metadata filtering. You possibly can run queries utilizing the AWS Administration Console or SDKs.

Put together a dataset to your Amazon Bedrock information base

On this publish, sample data set Describe a hypothetical online game and learn to use Amazon Bedrock’s information base to ingest and retrieve metadata. If you wish to proceed together with your AWS account, please obtain the file.

If you wish to add metadata to paperwork in an present information base, create a metadata file with the required file title and schema and proceed to Synchronize Information with a Information Base to begin incremental ingestion. .

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

title, description, genres, 12 months, writer, rating

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

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

Create a information base for Amazon Bedrock

For directions on creating a brand new information base, see Create a Information Base. This instance makes use of the next settings:

• in Organising the info supply backside of web page chunk techniquechoose No chunkingAs a result of we’ve got already preprocessed the doc within the earlier step.
• inside embedded mannequin part, choice Titan G1 Embed – Textual content.
• inside vector database part, choice Simply create a brand new vector retailer. Metadata filtering performance is obtainable for all supported vector shops.

Synchronize your dataset together with your information base

After you create your information base and have your information and metadata information in your Amazon Easy Storage Service (Amazon S3) bucket, you may start incremental ingestion. For directions, see Synchronize to convey information sources into your information base.

Querying with Metadata Filtering within the Amazon Bedrock Console

To make use of metadata filtering choices within the Amazon Bedrock console, observe these steps:

1. Within the Amazon Bedrock console, information base within the navigation pane.
2. Choose the information base you created.
3. select Take a look at information base.
4. please select composition Click on the icon to develop filter.
5. Enter the situation within the format key = worth (e.g. style = technique) and press . enter.
6. Choose a situation to alter the important thing, worth, or operator.
   
7. Proceed with remaining situations (e.g. (Style = Technique AND Yr >= 2023) OR (Score >= 9))
   
8. As soon as accomplished, enter your question within the message field and click on[走る。

この投稿では、「2023 年以降にリリースされるクールなグラフィックのストラテジー ゲーム」というクエリを入力します。

SDKを使用したメタデータフィルタリングによるクエリ

SDK を使用するには、まず Agents for Amazon Bedrock ランタイムのクライアントを作成します。

import boto3

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

次に、フィルターを構築します (以下はいくつかの例です)。

# genres = Strategy
single_filter= {
    "equals": {
        "key": "genres",
        "value": "Strategy"
    }
}

# genres = Strategy AND year >= 2023
one_group_filter= {
    "andAll": [
        {
            "equals": {
                "key": "genres",
                "value": "Strategy"
            }
        },
        {
            "GreaterThanOrEquals": {
                "key": "year",
                "value": 2023
            }
        }
    ]
} # (genres = Technique AND 12 months >=2023) OR rating >= 9 two_group_filter = { "orAll": [
        {
            "andAll": [
                {
                    "equals": {
                        "key": "genres",
                        "value": "Strategy"
                    }
                },
                {
                    "GreaterThanOrEquals": {
                        "key": "year",
                        "value": 2023
                    }
                }
            ]
        }, { "GreaterThanOrEquals": { "key": "rating", "worth": "9" } } ]}

move filter retrievalConfiguration Retrieve API or RetrieveAndGenerate API:

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

The next desk exhibits some responses with completely different metadata filtering situations.

Along with customized metadata, you can even filter utilizing S3 prefixes (that is built-in metadata, so that you need not present a metadata file). For instance, in case you manage your sport documentation into prefixes by writer, e.g. s3://$bucket_name/video_game/$writer/$game_id.csv), and might be filtered by particular publishers (e.g. neo_tokyo_games) use the next syntax:

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

cleansing

To wash up your sources, observe these steps:

1. Delete a information base.
   1. Within the Amazon Bedrock console, information base below orchestration within the navigation pane.
   2. Choose the information base you created.
   3. Be aware the next AWS Identification and Entry Administration (IAM) service position title. Information base overview part.
   4. inside vector database part, discover the gathering ARN.
   5. select erasethen sort “delete” and make sure.
2. Delete the vector database.
   1. Within the Amazon OpenSearch Service console, select: assortment below serverless within the navigation pane.
   2. Enter the gathering ARN you saved within the search bar.
   3. Chosen assortment chosen erase.
   4. On the affirmation immediate, sort “verify” and choose erase.
3. Delete an IAM service position.
   1. Within the IAM console, choose position within the navigation pane.
   2. Seek for the position title you famous earlier.
   3. Choose a task and choose erase.
   4. Enter the position title on the affirmation immediate to delete the position.
4. Delete the pattern dataset.
   1. Within the Amazon S3 console, navigate to the S3 bucket you used.
   2. Choose prefix and file, erase.
   3. If you wish to delete it, enter “Delete Completely” on the affirmation immediate.

conclusion

This publish described Amazon Bedrock’s information base metadata filtering function. You discovered the right way to add customized metadata to paperwork and use them as filters when retrieving and querying paperwork utilizing the Amazon Bedrock console and SDK. This improves the accuracy of the context and additional will increase the relevance of question responses whereas attaining a discount in the price of queries to vector databases.

See under for added sources.

Concerning the creator

Corvus Lee I’m a Senior Options Architect at GenAI Labs based mostly in London. He’s enthusiastic about designing and creating prototypes that use generative AI to resolve buyer issues. He additionally stays abreast of the newest developments by making use of generative AI and search strategies to real-world situations.

Ahmed Ewis is a Senior Options Architect at AWS GenAI Labs, the place he helps clients construct generative AI prototypes to resolve enterprise issues. After I’m not working with shoppers, I get pleasure from taking part in with my youngsters and cooking.

Chris Pecora Generative AI Information Scientist at Amazon Net Providers. He’s enthusiastic about constructing revolutionary merchandise and options, whereas additionally specializing in customer-focused science. When he isn’t conducting experiments or maintaining with the newest developments in GenAI, he loves spending time along with his youngsters.