The RAG System

The Core Idea: Process-Aware Indexing

The Dynamixs.AI RAG system does not only index relevant business content but also its state. Every indexed document carries the workflow status of the process instance it belongs to — and that status is automatically updated whenever the process moves forward.

This means you can query the knowledge base not just by content, but also by business context. When searching for contracts, for example, you can explicitly limit results to currently valid contracts — and exclude those that have been cancelled or superseded — simply by filtering on the workflow status at retrieval time.

Example: A contract document is indexed when it is approved. When the contract is later terminated, the process moves to a new status — and the RAG index is updated automatically in the background. A retrieval query looking for “valid contracts with supplier X” will no longer return the terminated contract, without any manual intervention.

Integration

Working with the RAG system involves two steps — the indexing phase where your business data is stored into the knowledge base, and the retrieving phase where relevant context is retrieved from your knowledge base at the right point in your process. Both are modelled directly in BPMN using dedicated adapter classes.

• org.imixs.ai.rag.workflow.RAGIndexPlugin — handles creating, updating, and deleting embeddings in the knowledge base
• org.imixs.ai.rag.workflow.RAGRetrievalAdapter — handles retrieval queries against the knowledge base

Before any indexing or retrieval can take place, the corresponding plugin class need to be registered in the Plugin List of your BPMN workflow model. Open the Workflow Properties of your process and add the plugin class org.imixs.ai.rag.workflow.RAGIndexPlugin:

Once registered, both classes are active for every processing cycle in that model. Their behaviour at each individual event is then controlled through the Workflow Result configuration of that event — as described in the sections below.

Indexing

The org.imixs.ai.rag.workflow.RAGIndexPlugin creates, updates, and deletes embeddings in the knowledge base. It is configured in the Workflow Result of a BPMN Event using name="INDEX":

<imixs-rag name="INDEX">
  <endpoint-embeddings>my-rag</endpoint-embeddings>
  <category>contracts</category>
  <debug>true</debug>
</imixs-rag>

The content that gets indexed is defined in a PromptDefinition Data Object associated with the same Event — exactly the same pattern used for LLM prompts. This means the full power of dynamic placeholders is available: <itemvalue> tags, <FILECONTEXT> for file attachments, and any other text adapter:

<PromptDefinition>
  <prompt>
    <itemvalue>$workflowgroup</itemvalue>: <itemvalue>$workflowstatus</itemvalue>
    <itemvalue>$workflowsummary</itemvalue>
    # Summary
    <itemvalue>$workflowabstract</itemvalue>
  </prompt>
</PromptDefinition>

At indexing time, the placeholders are resolved against the current process instance, and the resulting text is converted into a vector embedding and stored in the knowledge base together with the workflow metadata.

<imixs-rag name="DISABLE"></imixs-rag>

<imixs-rag name="PROMPT">
  <endpoint-completion>http://openai-api-completions/</endpoint-completion>
  <endpoint-embeddings>http://openai-api-embeddings/</endpoint-embeddings>
</imixs-rag>

<imixs-rag name="DELETE"></imixs-rag>

Retrieval

The org.imixs.ai.rag.workflow.RAGRetrievalAdapter searches the knowledge base for entries semantically similar to a given query and returns a list of matching process instance references. It is recommended to model RAG events in an asynchronous way (see the section Async Events).

The RAGRetrievalAdapter is configured in the Workflow Result of the corresponding BPMN Event using name="RETRIEVAL":

<imixs-rag name="RETRIEVAL">
  <endpoint-embeddings>http://openai-api-server/</endpoint-embeddings>
  <reference-item>product.ref</reference-item>
  <max-results>5</max-results>
  <modelgroups>Contract</modelgroups>
  <tasks>1200</tasks>
  <debug>true</debug>
</imixs-rag>

You can specify exactly which data should be used for the retrieval phase.

The modelgroups and tasks filters are where the process-awareness of the RAG system becomes directly useful in the model. By specifying a task ID that corresponds to the “active” or “valid” state of a process, you ensure that only currently relevant documents are considered — automatically excluding cancelled, expired, or superseded entries.

<PromptDefinition>
  <prompt><![CDATA[
    <itemvalue>request.subject</itemvalue>
    <FILECONTEXT>^.+\.eml$</FILECONTEXT>
  ]]></prompt>
</PromptDefinition>

Putting It Together

A typical RAG-enabled process in Dynamixs.AI works like this:

1. Index — When a contract is approved, an indexing event stores the contract content and metadata as a vector embedding, tagged with the current workflow status.
2. Auto-update — As the contract moves through its lifecycle (renewal, amendment, termination), the workflow metadata in the index is updated automatically.
3. Retrieve — When a new request arrives, a retrieval event searches the knowledge base for semantically relevant contracts — filtered to only those currently in an active status.
4. Generate — The retrieved contract references are passed to an LLM prompt as context, enabling the model to generate a response grounded in the actual, current contract data.