Add a reasoning field to a signature

with_reasoning() adds a string output field, reasoning by default, before the existing output fields, so the model writes out its reasoning before it answers (chain-of-thought prompting). chain_of_thought() builds a module from the result.

Usage

with_reasoning(
  x,
  prefix = "Let's think step by step in order to",
  reasoning_field = "reasoning",
  instructions = NULL,
  ...
)

Arguments

x

A signature from signature(), or a signature string such as "question -> answer".

prefix

Start of the reasoning field's description. The description reads "Reasoning: prefix produce the fields.", where fields names the original output fields.

reasoning_field

Name of the added field.

instructions

New instructions. With NULL (the default), existing instructions get " Think through your reasoning step by step before providing the answer." appended, and empty instructions are replaced by "Given inputs, think step by step and produce outputs."

...

Ignored.

Value

A new signature whose output is an object with the reasoning field first, followed by the original output fields. A bare output type becomes a field named answer.

See also

Other signatures: has_reasoning(), input(), signature(), with_instructions(), without_reasoning()

Examples

with_reasoning("question -> answer")
#> 
#> ── Signature ──
#> 
#> ── Inputs 
#> • question: "string" - Input: question
#> 
#> ── Output 
#> Type: "object(reasoning: string, answer: string)"
#> 
#> ── Instructions 
#> Given the fields `question`, produce the fields `answer`. Think through your
#> reasoning step by step before providing the answer.

# The prefix goes into the description of the reasoning field
sig <- with_reasoning(
  "math_problem -> solution: float",
  prefix = "Let me solve this step by step, then"
)
sig@output_type@properties$reasoning@description
#> [1] "Reasoning: Let me solve this step by step, then produce the solution."