DSPy Signature Generator

Name: dspy-signature
Rating: 92
Author: Qredence

Signatures define input/output structure for dspy.RLM tasks. Format: "input1, input2 -> output1, output2".

Signature Syntax

code

input_field1, input_field2 -> output_field1, output_field2

Rules:

•Use snake_case for field names
•Be descriptive: source_code not sc
•Use plural for lists: items, results
•Avoid Python reserved words: input, output, type
•No duplicate field names across inputs and outputs
•Must contain -> separator

Quick Examples

python

# Simple QA
"question -> answer"

# Context-based QA with confidence
"context, question -> answer, confidence"

# Code analysis
"code -> explanation, complexity_analysis"

# Document summarization
"document -> summary, key_points, word_count"

# RLM exploration
"dataset, query -> findings, statistics, insights"

For extensive examples by category, see references/signature-examples.md.

Usage with fleet-rlm

python

import dspy
from fleet_rlm import ModalInterpreter

signature = "question -> answer, confidence"

interpreter = ModalInterpreter(timeout=120)
rlm = dspy.RLM(
    signature=signature,
    interpreter=interpreter,
    max_iterations=5,
    max_llm_calls=10,
)

result = rlm(question="What is the capital of France?")
print(result.answer)       # Access via dot notation
print(result.confidence)   # NOT result["confidence"]

Common Field Types

Field	Description	Example Values
`text` / `document`	Raw or long-form content	String
`question` / `query`	Query to answer or search	String
`context`	Background information	String
`code`	Source code	String
`answer` / `result`	Single outcome	String
`results` / `items`	Multiple outcomes	List
`summary`	Condensed text	String
`count`	Numeric count	Integer
`confidence`	0-1 score	Float

Best Practices

DO:

•Be specific: python_code not code when context is clear
•Include metadata fields: confidence, explanation when useful
•Consider downstream consumers of the output
•Use consistent naming: same field name = same semantics

DON'T:

•Use single letters (x, q)
•Overload fields (one concept per field)
•Use dots in names (user.name -> user_name)

Advanced Patterns

Conditional Outputs

python

"code -> result, error, success"
# success: bool, result: output if True, error: message if False

Iteration Tracking

python

"task -> result, steps_taken, iterations_used"

fleet-rlm Built-in Signatures

Defined in src/fleet_rlm/signatures.py:

•ExtractArchitecture: docs, query -> modules, optimizers, design_principles
•ExtractAPIEndpoints: docs -> api_endpoints
•FindErrorPatterns: docs -> error_categories, total_errors_found
•ExtractWithCustomTool: docs -> headers, code_blocks, structure_summary
•AnalyzeLongDocument: document, query -> findings, answer, sections_examined
•SummarizeLongDocument: document, focus -> summary, key_points, coverage_pct
•ExtractFromLogs: logs, query -> matches, patterns, time_range