Vector Write and Read Nodes

To automate business processes effectively, your AI workflows require access to high-quality, well-structured, and reusable knowledge. You can use Vector Write and Read nodes in AI Agent Studio to store knowledge as embeddings and retrieve it intelligently with semantic search and metadata filters. By leveraging these nodes, you can design workflows that capture, manage, and reuse critical business knowledge across processes, thereby improving accuracy, scalability, and reliability in an enterprise environment.

Here's what you need to know to effectively use Vector Write and Read nodes in your workflow agents, along with some best practices for organizing and managing enterprise knowledge for optimal retrieval.

Vector Write Node

Vector write nodes store high-value knowledge as embeddings for future semantic retrieval. These nodes are also referred to as vector store nodes.

When you create a node, you can configure all values as expressions that resolve dynamically at runtime. Here's an example of the node creation window. Each numbered field is described in detail in the following table.

Vector write node creation window

Callout Number	Field	Description	Example
1	Name	Design-time name for the node	Use descriptive names, like `WriteResolutionSummaryToVectorDB`, so the workflow is easy to understand.
2	Code	Programmatic identifier in the workflow schema	This field is auto-generated but can be changed to a user-defined one, such as `write_resolution_vector`. Use lowercase and underscores.
3	Error Handler	Fallback path on failure	Select an error branch or a dedicated handler node to manage failures cleanly.
4	Operation Type	Method of how the document is written to the index	Use `INSERT` for new entries, `OVERWRITE`to replace content, `UPSERT` to safely update or create, or `DELETE` to remove entries.
5	Index Name	Name of the vector index to write to	Choose an existing index or specify a new one, such as `support_ticket_summaries` or `product_docs_index`.
6	Content	Textual data to embed	Summarize this in a clean and structured manner. Avoid raw logs. Instead, use LLM-generated summaries, extracted facts, or curated knowledge.
7	Content Type	Type of content being indexed or embedded	Typical values include `json` or `text`.
8	Document ID	Unique identifier for this record	Use stable identifiers, such as `ticket_1123` or `customer_450_profile`.
9	Properties	Additional optional metadata key-values	`{objectId:"a12345", region:“NA”, severity:“High”}`

Best Practices for Building Vector Write Nodes

A vector store works best when you write clear, meaningful, metadata-rich knowledge while avoiding noise, duplication, and topic drift.

Best Practice	Description
Store only public, reusable knowledge	Exclude sensitive, permission-controlled, or private information, for example, personally identifiable information (PII), credentials, confidential documents, or any information requiring permission checks.
Focus on durable, high-value content	Store information that will have long-term value for future workflows, such as case resolutions, structured summaries, or validated insights. Avoid writing noisy, one-off, or ephemeral content, like raw chat logs or temporary instructions.
Clean and normalize before writing	Always standardize content, summarize documents, deduplicate overlaps, remove irrelevant details, and attach key metadata.
Keep data up to date	Vector writes must reflect the latest truth. Stale entries decrease accuracy and lead to incorrect answers. Set auto-refresh triggers to update vectors whenever business objects change (for example, case closures, policy updates, PO resolutions). Use OVERWRITE or UPSERT to keep summaries up to date. Delete outdated or unused entries to prevent stale data from accumulating. Maintain freshness so workflows rely on accurate, up-to-date knowledge.
Follow smart update practices	Update vectors in a controlled and intentional manner. Don't create new entries for each update, or store conflicting or outdated data. Avoid creating new entries for every update. Refine existing data instead of duplicating. Use UPSERT to update and improve existing summaries. Use OVERWRITE for significant changes or new, higher-quality content. Prune redundant or conflicting entries to keep the vector store clean. Establish clear versioning strategies for policies, product versions, and configuration changes to prevent fragmentation.
Always include rich metadata	Tag with business object IDs, product numbers, version information, and other searchable attributes. Use a consistent metadata schema across agents and indexes.
Prevent data pollution	Check for existing knowledge to avoid duplicates and inconsistent tags. Regularly remove low-value content.
Exclude business object data	Duplicating this data in vector memory creates noise, redundancy, and version drift.

Vector Read Node

Vector read nodes retrieve the most relevant knowledge using semantic similarity and metadata filters.

Here's an overview of what the node creation window looks like. When creating a node, all values can be configured as expressions that will be resolved at runtime.

Vector read node creation window

Callout Number	Field	Description	Example
1	Name	Name of the node	Use clear names like `RetrieveTicketContextFromVectorDB` to keep the workflow readable.
2	Code	Internal programmatic identifier	This field is auto-generated but can be changed to a user-defined one, like `read_ticket_context_vector`.
3	Error Handler	Defines what occurs on failure	Route to an error handler or fallback logic to avoid empty or invalid agent responses.
4	Index Name	Vector index to search	Choose the same index used by the writer, such as `support_ticket_summaries` or `employee_profile_index`.
5	Query	Natural language search query	Avoid vague queries and ask for intent-specific info, such as, “What troubleshooting steps were taken?”
6	Document ID	Unique identifier for this record.	Specify this when you want details tied to a known record, such as `ticket_12345_summary`.
7	Data Fields	Metadata fields to return	Specify an array of strings that will be used in the filter.
8	Filter Criteria	Logical filters applied before ranking	Specify this field to enable high precision and to constrain results from a retrieval, such as `product = Payroll`, `region = US`, or `severity >= High`.
9	Maximum Results	Maximum number of ranked results returned	Specify an integer value for the max number of results to be displayed.

Best Practices for Building Vector Read Nodes

A vector read node works best when you write clear, specific queries that leverage metadata filters to ensure semantically accurate responses, while preventing topic drift and hallucinations.

Best Practice	Description
Write clear, intent-driven queries	Ensure queries are specific and aligned with the exact workflow or agent task. For example, use precise prompts like "What resolved similar issues?" instead of vague ones like "Help me with this."
Use metadata filters for precision	Apply filters (for example, entity type, business object ID, and product number) to target only relevant content.
Validate retrieved results before using them	Check metadata and context before trusting or using results in workflows. This validation prevents incorrect answers, silent workflow failures, and LLM hallucinations grounded in bad evidence.
Set an appropriate `maxResults` value	Set the `maxResults`field to return three to five, or more results. Don't depend on a single query result.
Add graceful fallback logic	Implement backup steps, such as business objects or API lookups, to handle cases where queries return no results.
Avoid broad or irrelevant retrieval	Scope queries to just what’s needed. Skip irrelevant or noisy indexes to boost performance and accuracy of retrieved results.