How Precisely transforms user experience with AI agents using Amazon S3 Vectors

At Precisely, the team is reimagining the user experience for its Data Integrity Suite by adding a conversational interface powered by AI agents to the traditional UI. With this enhancement, users can interact with the platform more naturally and intuitively (asking questions, making requests, and exploring data assets through dialogue) while still benefiting from the familiar features of a classic UI. The conversational interface enriches workflows, makes information discovery more accessible, and helps users express their needs in everyday language.

Amazon S3 Vectors—the first cloud object store with native support for storing and querying vectors—powers the solution by enabling robust similarity searches. It helps users quickly find relevant assets and insights by describing what they need, boosting efficiency and engagement. This post discusses how S3 Vectors is used in Precisely’s solution, explains why it is suitable, and covers additional relevant details. You will gain practical insight into architecting a conversational search experience on S3 Vectors, including tenant-aware indexing, metadata design, and cost modeling, so you can apply these patterns to your own multi-tenant applications.

Efficient, tenant-aware indexing with S3 Vectors

S3 Vectors offers a unique approach to vector storage and search that suits multi-tenant SaaS environments. One standout feature is the ability to maintain thousands of independent, lightweight vector indexes within a single S3 vector bucket. This flexibility lets organizations segment vectors by content type or, as Precisely does, create a separate index for each tenant. This design simplifies index management and maintains strong data isolation in multi-tenant environments, with no infrastructure to provision or maintain.

S3 Vectors accesses only the indexes involved in each query, so a tenant’s vectors enter the query path only when that tenant is active. That property shaped Precisely’s topology decision. Had all tenants been placed in a single shared index, even a sharded one, activity from a handful of tenants would pull unrelated tenants’ data into the query path, inflating resource usage and forcing the system to size for aggregate peak rather than actual demand. S3 Vectors also enforces request limits per index, so a tenant’s ingest burst or query spike is contained to that tenant’s index. Because indexes are serverless and billed on data stored and processed rather than on provisioned capacity, partitioning into thousands of small indexes costs no more than consolidating into one. That made it possible for Precisely to make physical isolation the default for every tenant rather than reserving it for the largest accounts, while scaling linearly with tenant count.

Using S3 Vectors for conversational search

Today, users expect to engage with data platforms like collaborating with colleagues. In Precisely’s conversational search agent, a user’s natural language request is first processed by a large language model (LLM), which interprets the query and generates a precise vector search phrase along with relevant metadata filters. These are used to initiate a similarity search using S3 Vectors. Users can describe their requests in everyday language and swiftly retrieve the most pertinent assets, insights, or responses. By shifting information discovery from a manual, menu-based process to an intuitive dialogue powered by LLM parsing and S3 vector similarity search, Precisely streamlines and accelerates access to desired information, regardless of the complexity or specificity of the inquiry.

S3 Vectors is well-suited for conversational search applications, delivering strong performance and consistently accurate results. The S3 Vectors query API takes the number of results to return, with no candidate-list depth or search-effort parameter to tune. S3 Vectors applies metadata filters as part of the query, so a filtered query remains likely to return a full set of relevant matches. This is especially valuable in conversational search scenarios, where both the search phrase and metadata filters are dynamically generated from user input, making it difficult to predict the optimal candidate count in advance. This reduces operational burden: with a separate index per tenant and tenant sizes spanning several orders of magnitude, there are no search parameters to calibrate per tenant or revisit as tenants grow. This enables applications to deliver a fast, accurate experience for users.

Intelligent recommendations: Powering proactive actions with S3 Vectors

Beyond simply responding to user queries, Precisely’s system uses S3 Vectors to proactively recommend actions that enhance productivity and data governance. For instance, during a conversation, the system might suggest discovering relevant data quality rules or curating metadata tailored to the user’s inquiry and the underlying data. These intelligent recommendations help users make informed decisions and integrate seamlessly into the natural workflow.

S3 Vectors enables intelligent recommendations by delivering consistent query performance, which is essential when generating proactive suggestions within conversational workflows. This responsiveness lets the system integrate vector-based recommendations with traditional strategies, such as surfacing linked assets from catalog relationships and prioritizing suggestions by quality scores, so users receive a diverse set of relevant actions and insights. Additionally, the architecture supports performing multiple vector searches in parallel. For example, it can identify direct recommendations and uncover related suggestions through similar items. This flexibility empowers the system to provide context-aware, timely recommendations that enhance productivity and data governance, all within the natural flow of user interaction.

Conversational search and intelligent recommendation workflows in the Precisely Data Integrity Suite

Conversational search in Data Integrity Suite spans two flows: an ingestion path that keeps the vector index current as the catalog changes, and a query path that serves user requests. Both rely on embeddings, which are numerical representations of content generated by an embedding model. Content with similar meaning produces vectors that sit close together, allowing a search to match meaning rather than exact wording. The ingestion path converts each catalog asset into a vector and stores it in an S3 Vectors index. The query path converts the user’s search phrase into a vector using the same model, and S3 Vectors returns the stored vectors closest to it.

Vector ingestion path in Precisely Data Integrity Suite

The vector ingestion path includes the following steps:

1. Changes to the catalog, whether made by users or by automated cataloging processes, are received by the Data Catalog Service in the Data Integrity Suite, which publishes a change event to Amazon Managed Streaming for Apache Kafka (Amazon MSK).
2. The Data Catalog Event Processor consumes these events.
3. The Event Processor vectorizes the affected assets using the Amazon Titan Text Embeddings V2 model available in Amazon Bedrock, producing a 512-dimension vector for each asset. Embeddings are generated in batch rather than one record at a time.

The resulting vectors, accompanied by relevant metadata for filtering purposes, are written to the tenant’s index using PutVectors, batched to increase ingest performance and optimize costs. Vector keys are derived from the catalog asset identifier, so re-embedding an edited asset overwrites the existing entry rather than creating a duplicate. S3 Vectors provides strong read-after-write consistency, so a newly cataloged asset becomes searchable on the next query with no index rebuild or refresh interval.

Query and recommendation paths in Precisely Data Integrity Suite

The query path consists of the following steps:

1. Upon receiving a user request through the conversational search interface, the request is routed to the Conversational Search Agent.
2. The user’s inquiry is forwarded to Amazon Bedrock, where an LLM analyzes and translates it into a sequence of actionable steps required to address the query.
3. If the LLM-generated steps necessitate a similarity search, the search phrase and applicable filters are employed to initiate the process through the Similarity Search Tool using Model Context Protocol (MCP).
4. The Similarity Search Tool invokes the Data Catalog Service to execute the similarity search.
5. The Data Catalog Service generates a vector representation of the search phrase using the Amazon Titan Text Embeddings V2 model in Amazon Bedrock.
6. This vector is used to query the tenant’s index with QueryVectors with the applicable metadata filters. The Conversational Search Agent receives query results from S3 Vectors and integrates them with other step results or uses them as input for further actions.
7. During this workflow, the Conversational Search Agent might identify opportunities for rule recommendations or metadata curation based on contextual information. When the agent detects such recommendation actions, it engages the Recommendation Tool.
8. The Recommendation Tool calls the Data Catalog Service to perform a context-based similarity search. The service generates the appropriate vector and runs the S3 Vectors query, then relays the results back to the Conversational Search Agent.

The following table summarizes the key parameter values for the solution.

Implementation details

In this section, we discuss the choices behind the solution workflow.

Index topology and scale

A single vector bucket holds up to 10,000 indexes, and each index holds up to 2 billion vectors, so one index per tenant scales with substantial headroom. Precisely anticipates tenants ranging from roughly 100,000 to 30 million catalog assets, and a single index absorbs that full range without any change to the topology. Beyond the per-bucket index limit, indexes shard across additional vector buckets with no change to the query path.

Metadata design

Each vector carries filterable metadata describing the asset. Type, class, catalog type, and data classification are common to all assets, with additional attributes specific to each asset type. This metadata also encodes what is needed to apply the platform’s access controls, so a similarity search is constrained to the assets the requesting user is permitted to see. Descriptive fields such as names, tags, and semantic types are carried as non-filterable metadata, returned alongside results but neither evaluated during the search nor counted toward the data processed in a query.

Access rights are stored as metadata values and applied as metadata filters at query time, so a similarity search only returns the assets the requesting user is permitted to see. A request for a given number of results never surfaces anything the user isn’t entitled to. Without metadata filtering, you would have to over-fetch top-K candidates and discard the ones the user can’t access, with no reliable way to size that over-fetch: a narrowly permissioned user might require retrieving many multiples of the requested count. Metadata filtering eliminates the problem entirely, removing the need for a tuning parameter on top-K requests.

Measured performance

Across three benchmark runs of 30,000 queries each, issued from 30 parallel processes against a single index containing 50 million vectors in the us-east-2 AWS Region, average query time was in the low hundreds of milliseconds, with the workload achieving hundreds of queries per second. These timings cover the S3 Vectors query call alone and exclude embedding generation. This benchmark corpus is larger than the biggest tenant Precisely anticipates, so these figures represent the upper end of the expected per-index scale. Ingestion runs asynchronously and is not on the user-facing path.

Cost-efficient operations and consistent accuracy with S3 Vectors

One of the key advantages of S3 Vectors is its cost-efficient operational model: query, load, and storage are all usage-based. Because charges track the logical size of the data stored and the volume of data processed per query rather than provisioned memory, Precisely stores full-precision embeddings and never had to evaluate quantization as a cost-reduction measure. The trade-off between memory footprint and result accuracy didn’t enter the design.

Precisely’s modeling confirms this. For a projected fleet of 87 tenants holding 80 million assets in total, with individual tenants ranging from 100,000 to 20 million assets, and even at a generous 87 million queries per month, the total cost is approximately $1,000 per month. Roughly three-quarters of that is query processing, with storage and the ingestion path accounting for the remainder. That query figure assumes every tenant issues the same volume regardless of size; scaling it with tenant size brings the total down further. Measured against the conversational search feature as a whole, vector storage and retrieval represent a small fraction of total run cost, which is dominated by LLM inference.

This approach keeps cost proportional to actual usage while maintaining search quality, making S3 Vectors a strong fit for this workload.

Conclusion

S3 Vectors stands out as a robust solution for modern conversational search, blending the intuitiveness of natural language processing with the speed and precision of vector-based retrieval. By simplifying search configuration, delivering consistent performance and accurate results, and enabling intelligent, context-aware recommendations, S3 Vectors transforms how users interact with and discover information in the Precisely Data Integrity Suite. Its usage-based operational model keeps cost proportional to actual usage and makes per-tenant index isolation practical at every tenant size without a cost premium. These benefits make S3 Vectors an ideal choice for empowering dynamic, user-friendly, and resource-conscious data platforms.

To learn more, see the Amazon S3 Vectors product page and the Amazon S3 Vectors documentation. To see how these capabilities are applied in practice, visit the Precisely Data Integrity Suite.