Understanding pipelines 

Common step types 

Query optimizer

Rewrites or clarifies the search query. Before retrieval it uses the current query and optional chat history. After retrieval it can also use the first retrieved context to decide whether a follow-up retrieval is needed. The shared JSON contract is:

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

Copied!

Before retrieval, query rewrites the search query and the retrieval decision normally remains all. After retrieval, one LLM run can decide whether the existing context is sufficient, create a follow-up query and select the next retrievers. none skips further retrieval, selected limits it to the named targets and all is the default. The response must use this JSON contract; configuration backwards compatibility does not mean that plain-text responses remain valid. The optimizer never generates the answer.

Retriever

Searches the vector store and stores retrieved documents in the context.

Context optimizer

Reduces, orders or restructures retrieved context before answer generation.

Answer generator

Generates the final user-facing answer from the current query, chat history, retrieved context and assistant rules.

Quality gate

Reviews the answer for grounding, consistency and rule compliance. Depending on configuration, it can pass, improve or reject an answer.

Example pipeline 

1. Query optimization before retrieval
2. Retrieval
3. Query decision after retrieval
4. Optional detail retrieval with the follow-up query
5. Context optimization
6. Answer generation
7. Quality gate

Not every project needs every step. For small sites, retrieval plus answer generation may be sufficient. For ambiguous queries, product data or legal-like content, query optimization and quality gates are usually helpful.

Validation and failures 

Before execution, the pipeline checks processor identifiers, step types and required predecessors. For example, a context optimizer needs an earlier retriever or memory step, and a quality gate needs an earlier answer generator. Invalid pipelines stop before the first processor runs.

stop aborts the pipeline when a step fails. continue records the failure and proceeds with the next step. The legacy value fallback behaves like continue and should no longer be selected for new configurations.

Only the final answer-producing step streams its output. This prevents a later quality gate from replacing text that has already been shown to the user.