Pipeline steps 

Common fields 

Title

Human-readable step name. Retriever-step titles must be unique within the pipeline because they also name the retrieval in prompts and stored context.

Stage

Position in the logical pipeline, such as pre-retrieval, retrieval, post-retrieval or answer generation.

Processor identifier

Registry identifier of the PHP processor service.

Prompt include flags

Decide whether identity, behaviour, retrieval and output rules from the assistant profile are included.

Step prompt fields

Step identity, behaviour, retrieval and output rules.

LLM settings

Model, temperature and max tokens.

Retrieval settings

optional connection and collection overrides, maximum retrieval results, score threshold, context limits and metadata fields. Each retriever adds a named group to the existing retrieval context.

Intent selector settings

An intent_selector step has an Intent selector outputs multi-select with retriever, presentation_rules and retriever_payload. Empty legacy values preserve the original retriever and presentation-rule behavior.

When retriever_payload is enabled, configure a dot-notated payload field such as meta.additional.category_ids. The AI returns only values in the retrieverPayload.values JSON array. The optional payload-retriever multi-select is populated from the current pipeline's retriever steps; an empty selection applies the filter to all retrievers. The payload value conjunction can be and or or and is applied by the server, not by the AI.

Post-retrieval query optimization

A query_optimizer step may be placed after retrieval. In that position it receives the current retrieval context and returns the shared decision contract:

{
    "query": "Which requirements apply to the application?",
    "retrieval": {
        "mode": "selected",
        "targets": ["detail_retrieval"]
    }
}

Copied!

none skips subsequent retrieval, selected runs only the named retriever steps and all remains the default. This combines the decision whether further information is needed with the follow-up query in one LLM run.

Failure strategy

Defines how the pipeline reacts if the step fails.

Response contract

LLM steps resolve a response contract through the registered contract services. The default contract produces plain text. intent_selector and query_optimizer use the decision contract, whose parts define retrieval selection, presentation-rule selection and optional retriever-payload values.

The same contract supplies the prompt instructions and parses the response. A step can explicitly select its decision contract through the contract field; otherwise the registry selects the contract supporting the processor type. Decision contracts use all, selected, all_except and none. For all and none, the target list must be empty.

Step prompt placement 

Prompts belong into assistant profile or pipeline step records. Processor code should stay generic and should not contain project-specific prompt text.