ruby_llm-contract 0.7.1 → 0.7.3

data/examples/00_basics.rb

@@ -1,500 +1,182 @@
 # frozen_string_literal: true
 
 # =============================================================================
-# EXAMPLE 0: From zero to ruby_llm-contract
+# EXAMPLE 0: SummarizeArticle — from plain prompt to full contract
 #
-# Starting from the simplest case — a plain string prompt —
-# and adding one layer at a time.
+# One step, seven incremental layers. Each layer adds exactly one capability
+# and shows the line of code that unlocks it. Start at Step 1, read top to
+# bottom, stop at the layer that matches your project.
 # =============================================================================
 
 require_relative "../lib/ruby_llm/contract"
 
-# Setup: test adapter returns canned responses (no real LLM needed)
-RubyLLM::Contract.configure do |config|
-  config.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"sentiment": "positive"}'
-  )
-end
-
-# =============================================================================
-# STEP 1: Simplest possible step — plain string prompt
-#
-# BEFORE (typical Rails code):
-#
-#   prompt = "Classify the sentiment of this text as positive, negative, or neutral. Return JSON."
-#   response = OpenAI::Client.new.chat(messages: [{role: "user", content: prompt + "\n\n" + text}])
-#   JSON.parse(response.dig("choices", 0, "message", "content"))
-#
-# Or with ruby_llm (one-liner, but still no validation):
-#
-#   RubyLLM.chat.ask("Classify the sentiment: #{text}")
-#
-# PROBLEM: no validation, no types, no trace, no structure
-# =============================================================================
-
-# Option A: with output_schema (recommended — simplest)
-class SimpleSentiment < RubyLLM::Contract::Step::Base
-  input_type String # plain Ruby class works!
-
-  output_schema do
-    string :sentiment
-  end
-
-  prompt do
-    user "Classify the sentiment of this text as positive, negative, or neutral. Return JSON.\n\n{input}"
-  end
-end
-
-result = SimpleSentiment.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive"}
-
-# Option B: with output_type (plain Ruby class — JSON parsing is implicit for Hash)
-class SimpleSentimentDryTypes < RubyLLM::Contract::Step::Base
-  input_type  String
-  output_type Hash
+ARTICLE = <<~ARTICLE
+  Ruby 3.4 ships with frozen string literals on by default, measurable YJIT
+  speedups on Rails workloads, and tightened Warning.warn category filtering.
+  Parser fixes and faster keyword argument handling land alongside.
+ARTICLE
+
+CANNED = {
+  tldr: "Ruby 3.4 brings frozen string literals by default, YJIT speedups, parser fixes.",
+  takeaways: [
+    "Frozen string literals are the default in Ruby 3.4",
+    "YJIT delivers measurable Rails speedups",
+    "Parser fixes and keyword argument handling improve"
+  ],
+  tone: "analytical"
+}.freeze
 
-  prompt do
-    user "Classify the sentiment of this text as positive, negative, or neutral. Return JSON.\n\n{input}"
-  end
-end
-
-result = SimpleSentimentDryTypes.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive"}
-
-# =============================================================================
-# STEP 2: Add system message — separate instructions from data
-#
-# BEFORE:
-#   Everything in one string — instructions and data mixed together
-#
-# AFTER:
-#   system = instructions (constant)
-#   user = data (variable)
-# =============================================================================
-
-class SentimentWithSystem < RubyLLM::Contract::Step::Base
-  input_type String
-
-  output_schema do
-    string :sentiment
-  end
-
-  prompt do
-    system "Classify the sentiment of the user's text."
-    user "{input}"
-  end
+RubyLLM::Contract.configure do |c|
+  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(response: CANNED)
 end
 
-result = SentimentWithSystem.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive"}
-
 # =============================================================================
-# STEP 3: Add rules — clear instructions for the model
-#
-# Rules are individual requirements. One rule per concern.
-# Much clearer than a single wall of text.
+# STEP 1 — Minimal: prompt + output_schema
+# The step enforces JSON shape. No business rules yet, no retry.
 # =============================================================================
 
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"sentiment": "positive", "confidence": 0.95}'
-  )
-end
+class SummarizeArticleMinimal < RubyLLM::Contract::Step::Base
+  prompt <<~PROMPT
+    Summarize this article for a UI card. Return a short TL;DR,
+    3 to 5 key takeaways, and a tone label.
 
-class SentimentWithRules < RubyLLM::Contract::Step::Base
-  input_type String
+    {input}
+  PROMPT
 
   output_schema do
-    string :sentiment, enum: %w[positive negative neutral]
-    number :confidence, minimum: 0.0, maximum: 1.0
-  end
-
-  prompt do
-    system "You are a sentiment classifier."
-    rule "Return JSON only."
-    rule "Use exactly one of: positive, negative, neutral."
-    rule "Include a confidence score from 0.0 to 1.0."
-    user "{input}"
+    string :tldr
+    array  :takeaways, of: :string, min_items: 3, max_items: 5
+    string :tone, enum: %w[neutral positive negative analytical]
   end
 end
 
-result = SentimentWithRules.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive", confidence: 0.95}
+r = SummarizeArticleMinimal.run(ARTICLE)
+r.status        # => :ok
+r.parsed_output # => {tldr: "...", takeaways: [...], tone: "analytical"}
 
 # =============================================================================
-# STEP 4: Add invariants — custom business logic on top of schema
-#
-# Schema handles structure (enums, ranges). Invariants handle logic
-# that schema can't express: conditional rules, cross-field checks, etc.
+# STEP 2 — Add a business rule (validate) that schema cannot express
+# Schema says "takeaways is an array of 3–5 strings". Nothing there says
+# "uniqueness" or "TL;DR fits the card". That is what validate blocks are for.
 # =============================================================================
 
-class SentimentValidated < RubyLLM::Contract::Step::Base
-  input_type String
-
-  output_schema do
-    string :sentiment, enum: %w[positive negative neutral]
-    number :confidence, minimum: 0.0, maximum: 1.0
-  end
-
-  prompt do
-    system "You are a sentiment classifier."
-    rule "Return JSON only."
-    rule "Use exactly one of: positive, negative, neutral."
-    rule "Include a confidence score from 0.0 to 1.0."
-    user "{input}"
-  end
-
-  # Schema already enforces enum + range. Validate adds custom logic:
-  validate("high confidence required for extreme sentiments") do |o|
-    next true unless %w[positive negative].include?(o[:sentiment])
-
-    o[:confidence] >= 0.7
-  end
+class SummarizeArticleValidated < SummarizeArticleMinimal
+  validate("TL;DR fits the card")  { |o, _| o[:tldr].length <= 200 }
+  validate("takeaways are unique") { |o, _| o[:takeaways].uniq.size == o[:takeaways].size }
 end
 
-# Happy path:
-result = SentimentValidated.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive", confidence: 0.95}
-
-# Model returns low confidence for extreme sentiment — invariant catches it:
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"sentiment": "positive", "confidence": 0.3}'
-  )
-end
-
-result = SentimentValidated.run("I love this product!")
-result.status            # => :validation_failed
-result.validation_errors # => ["high confidence required for extreme sentiments"]
-
-# Model returns non-JSON:
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: "I think it's positive"
-  )
-end
-
-result = SentimentValidated.run("I love this product!")
-result.status            # => :parse_error
-result.validation_errors # => ["Failed to parse JSON: ..."]
+r = SummarizeArticleValidated.run(ARTICLE)
+r.status             # => :ok
+r.validation_errors  # => []
 
 # =============================================================================
-# STEP 5: Add examples — show the model what you expect
-#
-# Few-shot: provide example input → output pairs.
-# The model better understands the expected format.
+# STEP 3 — Structured prompt (prompt AST: system, rule, section, user)
+# Replaces a heredoc. Individual nodes are reorderable, diffable, and
+# inspectable — useful when the prompt grows beyond a few lines.
 # =============================================================================
 
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"sentiment": "positive", "confidence": 0.92}'
-  )
-end
-
-class SentimentWithExample < RubyLLM::Contract::Step::Base
-  input_type String
-
-  output_schema do
-    string :sentiment, enum: %w[positive negative neutral]
-    number :confidence, minimum: 0.0, maximum: 1.0
-  end
-
+class SummarizeArticleStructured < RubyLLM::Contract::Step::Base
   prompt do
-    system "You are a sentiment classifier."
-    rule "Return JSON only."
-    rule "Use exactly one of: positive, negative, neutral."
-    rule "Include a confidence score from 0.0 to 1.0."
-    example input: "This is terrible", output: '{"sentiment": "negative", "confidence": 0.9}'
-    example input: "It works fine I guess", output: '{"sentiment": "neutral", "confidence": 0.6}'
-    user "{input}"
+    system "You summarize articles for a UI card."
+    rule   "Return valid JSON only."
+    rule   "Keep the TL;DR under 200 characters."
+    user   "{input}"
   end
-end
-
-result = SentimentWithExample.run("I love this product!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "positive", confidence: 0.92}
-
-# =============================================================================
-# STEP 6: Sections — replace heredoc string with structured AST
-#
-# BEFORE (typical heredoc prompt — one big string):
-#
-#   prompt = <<~PROMPT                                          # AFTER:
-#     You are a sentiment classifier for customer support.      # system "You are a sentiment classifier for customer support."
-#     Return JSON with sentiment, confidence, and reason.       # rule "Return JSON with sentiment, confidence, and reason."
-#                                                               #
-#     [CONTEXT]                                                 # section "CONTEXT",
-#     We sell software for freelancers.                         #   "We sell software for freelancers."
-#                                                               #
-#     [SCORING GUIDE]                                           # section "SCORING GUIDE",
-#     negative = complaint or frustration                       #   "negative = complaint or frustration\n
-#     positive = praise or thanks                               #    positive = praise or thanks\n
-#     neutral = question or factual statement                   #    neutral = question or factual statement"
-#                                                               #
-#     Classify this: #{text}                                    # user "Classify this: {input}"
-#   PROMPT                                                      #
-#
-# PROBLEM: one big string — can't reorder, diff, or reuse individual sections
-# AFTER: each part is a separate node in the prompt AST
-# =============================================================================
-
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"sentiment": "negative", "confidence": 0.85, "reason": "product complaint"}'
-  )
-end
-
-class SentimentWithSections < RubyLLM::Contract::Step::Base
-  input_type String
 
   output_schema do
-    string :sentiment, enum: %w[positive negative neutral]
-    number :confidence, minimum: 0.0, maximum: 1.0
-    string :reason
-  end
-
-  prompt do
-    system "You are a sentiment classifier for customer support."
-    rule "Return JSON with sentiment, confidence, and reason."
-
-    section "CONTEXT", "We sell software for freelancers."
-    section "SCORING GUIDE",
-            "negative = complaint or frustration\npositive = praise or thanks\nneutral = question or factual statement"
-
-    user "Classify this: {input}"
+    string :tldr
+    array  :takeaways, of: :string, min_items: 3, max_items: 5
+    string :tone, enum: %w[neutral positive negative analytical]
   end
 end
 
-result = SentimentWithSections.run("Your billing page is broken again!")
-result.status        # => :ok
-result.parsed_output # => {sentiment: "negative", confidence: 0.85, reason: "product complaint"}
-
 # =============================================================================
-# STEP 7: Hash input — multiple fields with auto-interpolation
-#
-# When input is a Hash, each key becomes a template variable.
-# {title} resolves to input[:title], {language} to input[:language], etc.
-# No manual string building needed.
+# STEP 4 — Hash input with variable interpolation
+# When you need more than raw text (audience, language, tenant), take a Hash
+# and reference its keys directly in the prompt.
 # =============================================================================
 
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"category": "billing", "priority": "high"}'
-  )
-end
-
-class ClassifyTicket < RubyLLM::Contract::Step::Base
+class SummarizeArticleMultiField < RubyLLM::Contract::Step::Base
   input_type RubyLLM::Contract::Types::Hash.schema(
-    title: RubyLLM::Contract::Types::String,
-    body: RubyLLM::Contract::Types::String,
+    article:  RubyLLM::Contract::Types::String,
+    audience: RubyLLM::Contract::Types::String,
     language: RubyLLM::Contract::Types::String
   )
 
-  output_schema do
-    string :category, enum: %w[billing technical feature_request other]
-    string :priority, enum: %w[low medium high urgent]
+  prompt do
+    system  "You summarize articles for a UI card."
+    rule    "Write the TL;DR and takeaways in {language}."
+    section "AUDIENCE", "{audience}"
+    user    "{article}"
   end
 
-  prompt do
-    system "You classify customer support tickets."
-    rule "Return JSON with category and priority."
-    rule "Respond in {language}."
-    user "Title: {title}\n\nBody: {body}"
+  output_schema do
+    string :tldr
+    array  :takeaways, of: :string, min_items: 3, max_items: 5
+    string :tone, enum: %w[neutral positive negative analytical]
   end
 end
 
-result = ClassifyTicket.run(
-  { title: "Can't update credit card", body: "Payment page gives error 500", language: "en" }
-)
-result.status        # => :ok
-result.parsed_output # => {category: "billing", priority: "high"}
-
 # =============================================================================
-# STEP 8: 2-arity invariants — validate output against input
-#
-# Sometimes you need to check that the output is consistent with the input.
-# A 2-arity invariant receives both |output, input| so you can cross-validate.
+# STEP 5 — 2-arity validate: check the output against the input
+# Catches "lazy" models that echo the article verbatim into the TL;DR.
+# The block receives |output, input| — pass the input-side check too.
 # =============================================================================
 
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"translation": "Bonjour le monde", "source_lang": "en", "target_lang": "fr"}'
-  )
-end
-
-class Translate < RubyLLM::Contract::Step::Base
-  input_type RubyLLM::Contract::Types::Hash.schema(
-    text: RubyLLM::Contract::Types::String,
-    target_lang: RubyLLM::Contract::Types::String
-  )
-
-  output_schema do
-    string :translation, min_length: 1
-    string :source_lang
-    string :target_lang
-  end
-
-  prompt do
-    system "Translate the text to the target language."
-    rule "Return JSON with translation, source_lang, and target_lang."
-    user "Translate to {target_lang}: {text}"
-  end
-
-  # Schema handles: translation non-empty, all fields present
-  # 2-arity validate: cross-validate output against input
-  validate("target_lang must match requested language") do |output, input|
-    output[:target_lang] == input[:target_lang]
+class SummarizeArticleFaithful < SummarizeArticleValidated
+  validate("TL;DR is shorter than the article") do |output, input|
+    output[:tldr].length < input.length / 2
   end
 end
 
-result = Translate.run({ text: "Hello world", target_lang: "fr" })
-result.status        # => :ok
-result.parsed_output # => {translation: "Bonjour le monde", source_lang: "en", target_lang: "fr"}
-
-# What if model returns wrong target language?
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"translation": "Hola mundo", "source_lang": "en", "target_lang": "es"}'
-  )
-end
-
-result = Translate.run({ text: "Hello world", target_lang: "fr" })
-result.status            # => :validation_failed
-result.validation_errors # => ["target_lang must match requested language"]
-
 # =============================================================================
-# STEP 9: Context override — per-run adapter and model
-#
-# Global config sets defaults. You can override per call via context.
-# Useful for: testing, switching models, A/B testing prompts.
+# STEP 6 — Retry with model fallback
+# Start on the cheapest model. If validate or schema rejects the output,
+# the gem automatically retries on the next model in the list.
 # =============================================================================
 
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(response: '{"sentiment": "positive"}')
-  c.default_model = "gpt-4.1-mini"
+class SummarizeArticleWithRetry < SummarizeArticleValidated
+  retry_policy models: %w[gpt-5-nano gpt-5-mini gpt-5]
 end
 
-# Uses global defaults:
-result = SimpleSentiment.run("I love this product!")
-result.status        # => :ok
-result.trace[:model] # => "gpt-4.1-mini"
-
-# Override adapter and model for this specific call:
-other_adapter = RubyLLM::Contract::Adapters::Test.new(response: '{"sentiment": "neutral"}')
-result = SimpleSentiment.run("I love this product!", context: { adapter: other_adapter, model: "gpt-5" })
-result.status          # => :ok
-result.parsed_output   # => {sentiment: "neutral"}
-result.trace[:model]   # => "gpt-5"
-
 # =============================================================================
-# STEP 10: StepResult — everything you get back from a run
-#
-# Every .run() returns a StepResult with status, output, errors, and trace.
+# STEP 7 — Inspect the Result: status, parsed_output, trace, per-attempt
+# Every run returns a value object with everything you need to log, debug,
+# or surface in an admin UI.
 # =============================================================================
 
-adapter = RubyLLM::Contract::Adapters::Test.new(response: '{"sentiment": "positive", "confidence": 0.92}')
-result = SentimentValidated.run("I love this product!", context: { adapter: adapter, model: "gpt-4.1-mini" })
-
-result.status            # => :ok
-result.ok?               # => true
-result.failed?           # => false
-result.raw_output        # => '{"sentiment": "positive", "confidence": 0.92}'
-result.parsed_output     # => {sentiment: "positive", confidence: 0.92}
-result.validation_errors # => []
-result.trace[:model]     # => "gpt-4.1-mini"
-result.trace[:latency_ms] # => 0     (instant with test adapter)
-result.trace[:messages] # => [{role: :system, content: "..."}, {role: :user, content: "..."}]
-
-# On failure, you still get everything for debugging:
-bad_adapter = RubyLLM::Contract::Adapters::Test.new(response: '{"sentiment": "positive", "confidence": 0.1}')
-result = SentimentValidated.run("I love this product!", context: { adapter: bad_adapter })
-
-result.status            # => :validation_failed
-result.ok?               # => false
-result.failed?           # => true
-result.raw_output        # => '{"sentiment": "positive", "confidence": 0.1}'
-result.parsed_output     # => {sentiment: "positive", confidence: 0.1}
-result.validation_errors # => ["high confidence required for extreme sentiments"]
+r = SummarizeArticleWithRetry.run(ARTICLE)
+r.status             # => :ok
+r.ok?                # => true
+r.parsed_output      # => {tldr: "...", takeaways: [...], tone: "analytical"}
+r.validation_errors  # => []
+r.trace[:model]      # => "gpt-5-nano"   (first model that passed)
+r.trace[:attempts]   # => [{attempt: 1, model: "gpt-5-nano", status: :ok, ...}]
+r.trace[:cost]       # => sum of per-attempt costs
 
 # =============================================================================
-# STEP 11: Pipeline — chain multiple steps with fail-fast
+# STEP 8 — Swap the Test adapter for a real LLM
+# The step itself does not change. Point ruby_llm at your provider and
+# pass Adapters::RubyLLM.new in context.
 #
-# Pipeline::Base composes steps into a sequence.
-# Output of step N automatically becomes input to step N+1.
-# If any step fails, execution halts immediately.
+# See examples/02_real_llm_minimal.rb for a runnable ~30-line version.
 # =============================================================================
 
-# Step A: classify sentiment
-class PipelineSentiment < RubyLLM::Contract::Step::Base
-  input_type String
-
-  output_schema do
-    string :text
-    string :sentiment, enum: %w[positive negative neutral]
-  end
-
-  prompt do
-    system "Classify sentiment and return the original text."
-    user "{input}"
-  end
-end
-
-# Step B: generate a response based on sentiment
-class PipelineRespond < RubyLLM::Contract::Step::Base
-  input_type Hash
-
-  output_schema do
-    string :response
-    string :tone
-  end
-
-  prompt do
-    system "Generate a customer support response matching the sentiment."
-    user "Text: {text}\nSentiment: {sentiment}"
-  end
-end
-
-# Pipeline: sentiment → respond
-class SupportPipeline < RubyLLM::Contract::Pipeline::Base
-  step PipelineSentiment, as: :classify
-  step PipelineRespond,   as: :respond
-end
-
-# Happy path:
-RubyLLM::Contract.configure do |c|
-  c.default_adapter = RubyLLM::Contract::Adapters::Test.new(
-    response: '{"text": "I love this product!", "sentiment": "positive"}'
-  )
-end
-
-# NOTE: with Test adapter, both steps get the same canned response.
-# With a real LLM, each step would get a different response.
-result = SupportPipeline.run("I love this product!")
-result.ok?                            # => true
-result.outputs_by_step[:classify]     # => {text: "I love this product!", sentiment: "positive"}
-result.outputs_by_step[:respond]      # => {text: "I love this product!", sentiment: "positive"}
-result.step_results.length            # => 2
+#   RubyLLM.configure { |c| c.openai_api_key = ENV.fetch("OPENAI_API_KEY") }
+#   adapter = RubyLLM::Contract::Adapters::RubyLLM.new
+#   result  = SummarizeArticleWithRetry.run(article_text, context: { adapter: adapter })
+#
+# Switch provider per call — ruby_llm resolves the provider from the model name:
+#   SummarizeArticleWithRetry.run(article_text, context: { adapter: adapter, model: "claude-sonnet-4-6" })
+#   SummarizeArticleWithRetry.run(article_text, context: { adapter: adapter, model: "gemma3:4b" })  # local Ollama
 
 # =============================================================================
-# SUMMARY
-#
-# Step 1:  user "{input}"                   — plain string, nothing else
-# Step 2:  system + user                    — separate instructions from data
-# Step 3:  + output_schema                  — declarative output structure
-# Step 4:  + invariants                     — custom business logic on top
-# Step 5:  + examples                       — few-shot
-# Step 6:  + sections                       — labeled context blocks
-# Step 7:  Hash input                       — multiple fields, auto-interpolation
-# Step 8:  2-arity invariants               — cross-validate output vs input
-# Step 9:  context override                 — per-run adapter and model
-# Step 10: StepResult                       — full status, output, errors, trace
-# Step 11: Pipeline                         — chain steps with fail-fast
+# Where to go next
 #
-# Each step adds one layer. Use as many as you need.
-# Even Step 1 gives you: typed input, JSON parsing, and trace.
+# 01_fallback_showcase.rb       — see the retry loop run in 30 seconds
+# 02_real_llm_minimal.rb        — swap Test adapter for Adapters::RubyLLM
+# 03_summarize_with_keywords.rb — growing prompt: add a keywords field
+# 04_summarize_and_translate.rb — pipeline: summarize → translate → review
+# 05_eval_dataset.rb            — define_eval, add_case, regression detection
+# 06_retry_variants.rb          — attempts: 3, reasoning_effort, cross-provider
 # =============================================================================