Step-back prompting

step_back() turns a specific question into a conceptually broader one. It's a query-transformation technique for RAG: the more general question retrieves broad-context documents (principles, categories, fundamentals) that searching the original question alone tends to miss. The pattern is to search with both and merge the passages.

from jangada_ai import LLM, step_back

llm = LLM("openai", "gpt-4o-mini")
broader = step_back(llm, "What are the treatment options for cataracts?")
print(broader)
# "What are the surgical and pharmacological approaches for lens opacity
#  management?"

It's just LLM + structured output, so it works on any provider — it doesn't depend on vision or the [rag] extra.

Typical use with RAG

Search the context with the original question and the step-back one, then merge the results before answering:

from jangada_ai import LLM, step_back
from jangada_ai.rag import RAG, vector_store

llm = LLM("openai", "gpt-4o-mini")
emb = LLM("openai", "text-embedding-3-small")
rag = RAG(emb, vector_store("postgresql://..."), chat=llm)

question = "What are the treatment options for cataracts?"
broader = step_back(llm, question)

specific = rag.search(question)
general = rag.search(broader)
# combine `specific` + `general` (dedupe) and answer with the merged context.

Parameters

step_back(
    llm,
    query,                          # the specific question
    instructions="Medical domain; answer in English.",  # ADDS to the default prompt
    prompt=None,                    # overrides the whole instruction (optional)
    params={"temperature": 0.2},   # generation params (optional)
)

• instructions is added to the default prompt — useful to pin the domain, language or what to emphasize, without losing the schema-guaranteed format.
• prompt replaces the whole instruction (the schema still guarantees the output). When overriding, include the question in your text — {query} is only interpolated in the default prompt.

Async version: await astep_back(llm, query, ...).

Robustness

Parsing is tolerant: it uses the structured output when valid, falls back to the raw text if the model ignores the schema and, as a last resort, returns the query itself — it never returns empty.

See also RAG and Structured output.