synth-ai 0.2.8.dev4__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

examples/blog_posts/mipro/configs/banking77_mipro_local.toml

@@ -0,0 +1,91 @@
+# MIPROv2 Prompt Learning for Banking77
+# Local backend configuration targeting the Banking77 intent classification task app.
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-web-web.modal.run"
+task_app_id = "banking77"
+
+# Initial prompt pattern (pattern-based mode)
+[prompt_learning.initial_prompt]
+id = "banking77_pattern"
+name = "Banking77 Classification Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "You are an expert banking assistant that classifies customer queries into banking intents. Given a customer message, respond with exactly one intent label from the provided list using the `banking77_classify` tool."
+order = 0
+
+[[prompt_learning.initial_prompt.messages]]
+role = "user"
+pattern = "Customer Query: {query}\n\nClassify this query into one of the banking intents using the tool call."
+order = 1
+
+[prompt_learning.initial_prompt.wildcards]
+query = "REQUIRED"  # Will be provided by task app at runtime
+
+# Policy configuration (model, provider, etc.)
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "openai/gpt-oss-20b"
+provider = "groq"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-mipro"  # Required for Banking77 task app
+
+# Training split config
+[prompt_learning.env_config]
+pool = "train"
+
+# MIPROv2-specific configuration
+[prompt_learning.mipro]
+env_name = "banking77"
+num_iterations = 5  # 5 iterations × 2 per iteration = 10 total prompt combos
+num_evaluations_per_iteration = 2  # 2 proposals per iteration
+batch_size = 6
+max_concurrent = 16
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.85
+max_instructions = 3
+
+# Retry guard for duplicate proposals/candidates
+duplicate_retry_limit = 10
+
+# Seed pools for different phases - tons of data for reliable evaluation
+bootstrap_train_seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]  # 15 seeds for bootstrap
+online_pool = [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]  # 25 seeds for online evaluation
+test_pool = [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]  # 10 held-out seeds for final evaluation
+
+# TPE hyperparameters
+[prompt_learning.mipro.tpe]
+gamma = 0.25              # Quantile for splitting good/bad observations
+n_candidates = 32         # Number of candidates to evaluate Expected Improvement
+n_startup_trials = 10     # Uniform random trials before TPE kicks in
+epsilon = 0.1             # Exploration probability (uniform sample)
+categorical_alpha = 1.0   # Laplace smoothing for categorical variables
+prior_weight = 0.05       # Prior mixing weight to avoid zero-densities
+
+# Demo set selection hyperparameters
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 5  # Maximum number of few-shot examples to include
+sets_per_size = 6          # Number of demo sets to generate per size
+include_empty = true       # Whether to include empty demo set (zero-shot)
+
+# Grounding configuration for instruction proposals
+[prompt_learning.mipro.grounding]
+n = 10                     # Number of instruction proposals to generate
+temperature = 0.7          # Temperature for LLM proposal generation
+max_tokens = 600           # Max tokens for proposal generation
+
+# Meta-update configuration
+[prompt_learning.mipro.meta_update]
+enabled = true             # Enable periodic meta-updates
+every_iterations = 3       # Regenerate instructions every N iterations
+topk_success = 5           # Number of top success examples to use for grounding
+topk_failure = 5           # Number of top failure examples to use for grounding
+validate_on_batch = 32     # Batch size for proposal validation
+keep_k = 12                # Maximum number of instruction variants to keep
+dedup_token_overlap = 0.8  # Token overlap threshold for deduplication
+regen_temperature_decay = 0.95  # Temperature decay factor per meta-update
+

examples/blog_posts/mipro/configs/banking77_mipro_test.toml

@@ -0,0 +1,87 @@
+# MIPROv2 Prompt Learning for Banking77
+# Test configuration with reduced iterations for faster testing
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-web-web.modal.run"
+task_app_id = "banking77"
+
+# Initial prompt pattern (pattern-based mode)
+[prompt_learning.initial_prompt]
+id = "banking77_pattern"
+name = "Banking77 Classification Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "You are an expert banking assistant that classifies customer queries into banking intents. Given a customer message, respond with exactly one intent label from the provided list using the `banking77_classify` tool."
+order = 0
+
+[[prompt_learning.initial_prompt.messages]]
+role = "user"
+pattern = "Customer Query: {query}\n\nClassify this query into one of the banking intents using the tool call."
+order = 1
+
+[prompt_learning.initial_prompt.wildcards]
+query = "REQUIRED"
+
+# Policy configuration
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "openai/gpt-oss-20b"
+provider = "groq"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-mipro"
+
+# Training split config
+[prompt_learning.env_config]
+pool = "train"
+
+# MIPROv2-specific configuration (reduced for testing)
+[prompt_learning.mipro]
+env_name = "banking77"
+num_iterations = 10  # Reduced for testing (default: 16)
+num_evaluations_per_iteration = 4  # Reduced for testing (default: 6)
+batch_size = 4
+max_concurrent = 10
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.80  # Lower threshold for testing (default: 0.85)
+
+# Seed pools for different phases
+bootstrap_train_seeds = [0, 1, 2, 3, 4]  # 5 seeds for bootstrap phase
+online_pool = [5, 6, 7, 8, 9]  # 5 seeds for online evaluation
+test_pool = [10, 11, 12, 13, 14]  # 5 seeds for final test
+
+# TPE hyperparameters
+[prompt_learning.mipro.tpe]
+gamma = 0.25              # Quantile for splitting good/bad observations
+n_candidates = 16         # Number of candidates to evaluate Expected Improvement (reduced for testing)
+n_startup_trials = 5      # Uniform random trials before TPE kicks in
+epsilon = 0.1             # Exploration probability (uniform sample)
+categorical_alpha = 1.0   # Laplace smoothing for categorical variables
+prior_weight = 0.05       # Prior mixing weight to avoid zero-densities
+
+# Demo set selection hyperparameters
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 5  # Maximum number of few-shot examples to include
+sets_per_size = 3          # Number of demo sets to generate per size (reduced for testing)
+include_empty = true       # Whether to include empty demo set (zero-shot)
+
+# Grounding configuration for instruction proposals
+[prompt_learning.mipro.grounding]
+n = 8                      # Number of instruction proposals to generate (reduced for testing)
+temperature = 0.7          # Temperature for LLM proposal generation
+max_tokens = 600           # Max tokens for proposal generation
+
+# Meta-update configuration
+[prompt_learning.mipro.meta_update]
+enabled = true             # Enable periodic meta-updates
+every_iterations = 3       # Regenerate instructions every N iterations
+topk_success = 3           # Number of top success examples to use for grounding (reduced for testing)
+topk_failure = 3           # Number of top failure examples to use for grounding (reduced for testing)
+validate_on_batch = 16     # Batch size for proposal validation (reduced for testing)
+keep_k = 8                 # Maximum number of instruction variants to keep (reduced for testing)
+dedup_token_overlap = 0.8  # Token overlap threshold for deduplication
+regen_temperature_decay = 0.95  # Temperature decay factor per meta-update
+

examples/blog_posts/mipro/configs/banking77_pipeline_mipro_gemini_flash_lite_local.toml

@@ -0,0 +1,98 @@
+# MIPROv2 Prompt Learning for the Banking77 Two-Step Pipeline
+# Using gemini-2.5-flash-lite as the policy model
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-pipeline-web-web.modal.run"
+task_app_id = "banking77-pipeline"
+
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_pattern"
+name = "Banking77 Pipeline Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "Pipeline placeholder message. Actual classifier/calibrator instructions are provided via metadata.pipeline_modules."
+order = 0
+
+[prompt_learning.initial_prompt.metadata]
+pipeline_modules = [
+  { name = "classifier", instruction_text = "You are an expert banking assistant. Given a customer query, choose the correct Banking77 intent and call the banking77_classify tool with that label.", few_shots = [] },
+  { name = "calibrator", instruction_text = "You review the classifier's suggestion. Confirm it if appropriate; otherwise pick the best Banking77 intent and return it via the banking77_classify tool.", few_shots = [] }
+]
+
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "gemini-2.5-flash-lite"
+provider = "google"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-pipeline"
+
+[prompt_learning.env_config]
+pool = "train"
+
+[prompt_learning.mipro]
+env_name = "banking77_pipeline"
+num_iterations = 5
+num_evaluations_per_iteration = 2
+batch_size = 6
+max_concurrent = 16
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.85
+max_instructions = 3
+duplicate_retry_limit = 10
+bootstrap_train_seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
+online_pool = [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]
+test_pool = [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]
+# Reference pool: Used to generate rich dataset context (~50k tokens) in the meta-prompt
+# Must not overlap with bootstrap/online/test pools to avoid data leakage
+reference_pool = [50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149]
+
+# System specification (pipeline-level, covers both analyzer and classifier stages)
+spec_path = "examples/task_apps/banking77_pipeline/banking77_pipeline_spec.json"
+spec_max_tokens = 5000
+spec_include_examples = true
+spec_priority_threshold = 8  # Only include high-priority rules (8+)
+
+[prompt_learning.mipro.tpe]
+gamma = 0.25
+n_candidates = 32
+n_startup_trials = 10
+epsilon = 0.1
+categorical_alpha = 1.0
+prior_weight = 0.05
+
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 5
+sets_per_size = 6
+include_empty = true
+
+[prompt_learning.mipro.grounding]
+n = 10
+temperature = 0.7
+max_tokens = 600
+
+[prompt_learning.mipro.meta_update]
+enabled = true
+every_iterations = 3
+topk_success = 5
+topk_failure = 5
+validate_on_batch = 32
+keep_k = 12
+dedup_token_overlap = 0.8
+regen_temperature_decay = 0.95
+
+[[prompt_learning.mipro.modules]]
+module_id = "classifier"
+max_instruction_slots = 3
+max_demo_slots = 5
+
+[[prompt_learning.mipro.modules]]
+module_id = "calibrator"
+max_instruction_slots = 3
+max_demo_slots = 5
+
+
+

examples/blog_posts/mipro/configs/banking77_pipeline_mipro_gpt41mini_local.toml

@@ -0,0 +1,96 @@
+# MIPROv2 Prompt Learning for the Banking77 Two-Step Pipeline
+# Using gpt-4.1-mini as the policy model
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-pipeline-web-web.modal.run"
+task_app_id = "banking77-pipeline"
+
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_pattern"
+name = "Banking77 Pipeline Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "Pipeline placeholder message. Actual classifier/calibrator instructions are provided via metadata.pipeline_modules."
+order = 0
+
+[prompt_learning.initial_prompt.metadata]
+pipeline_modules = [
+  { name = "classifier", instruction_text = "You are an expert banking assistant. Given a customer query, choose the correct Banking77 intent and call the banking77_classify tool with that label.", few_shots = [] },
+  { name = "calibrator", instruction_text = "You review the classifier's suggestion. Confirm it if appropriate; otherwise pick the best Banking77 intent and return it via the banking77_classify tool.", few_shots = [] }
+]
+
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "gpt-4.1-mini"
+provider = "openai"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-pipeline"
+
+[prompt_learning.env_config]
+pool = "train"
+
+[prompt_learning.mipro]
+env_name = "banking77_pipeline"
+num_iterations = 5
+num_evaluations_per_iteration = 2
+batch_size = 6
+max_concurrent = 16
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.85
+max_instructions = 3
+duplicate_retry_limit = 10
+bootstrap_train_seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
+online_pool = [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]
+test_pool = [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]
+# Reference pool: Used to generate rich dataset context (~50k tokens) in the meta-prompt
+# Must not overlap with bootstrap/online/test pools to avoid data leakage
+reference_pool = [50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149]
+
+# System specification (pipeline-level, covers both analyzer and classifier stages)
+spec_path = "examples/task_apps/banking77_pipeline/banking77_pipeline_spec.json"
+spec_max_tokens = 5000
+spec_include_examples = true
+spec_priority_threshold = 8  # Only include high-priority rules (8+)
+
+[prompt_learning.mipro.tpe]
+gamma = 0.25
+n_candidates = 32
+n_startup_trials = 10
+epsilon = 0.1
+categorical_alpha = 1.0
+prior_weight = 0.05
+
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 5
+sets_per_size = 6
+include_empty = true
+
+[prompt_learning.mipro.grounding]
+n = 10
+temperature = 0.7
+max_tokens = 600
+
+[prompt_learning.mipro.meta_update]
+enabled = true
+every_iterations = 3
+topk_success = 5
+topk_failure = 5
+validate_on_batch = 32
+keep_k = 12
+dedup_token_overlap = 0.8
+regen_temperature_decay = 0.95
+
+[[prompt_learning.mipro.modules]]
+module_id = "classifier"
+max_instruction_slots = 3
+max_demo_slots = 5
+
+[[prompt_learning.mipro.modules]]
+module_id = "calibrator"
+max_instruction_slots = 3
+max_demo_slots = 5
+

examples/blog_posts/mipro/configs/banking77_pipeline_mipro_local.toml

@@ -0,0 +1,94 @@
+# MIPROv2 Prompt Learning for the Banking77 Two-Step Pipeline
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-pipeline-web-web.modal.run"
+task_app_id = "banking77-pipeline"
+
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_pattern"
+name = "Banking77 Pipeline Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "Pipeline placeholder message. Actual classifier/calibrator instructions are provided via metadata.pipeline_modules."
+order = 0
+
+[prompt_learning.initial_prompt.metadata]
+pipeline_modules = [
+  { name = "classifier", instruction_text = "You are an expert banking assistant. Given a customer query, choose the correct Banking77 intent and call the banking77_classify tool with that label.", few_shots = [] },
+  { name = "calibrator", instruction_text = "You review the classifier's suggestion. Confirm it if appropriate; otherwise pick the best Banking77 intent and return it via the banking77_classify tool.", few_shots = [] }
+]
+
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "openai/gpt-oss-120b"
+provider = "groq"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-pipeline"
+
+[prompt_learning.env_config]
+pool = "train"
+
+[prompt_learning.mipro]
+env_name = "banking77_pipeline"
+num_iterations = 5
+num_evaluations_per_iteration = 2
+batch_size = 6
+max_concurrent = 16
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.85
+max_instructions = 3
+duplicate_retry_limit = 10
+bootstrap_train_seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
+online_pool = [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]
+test_pool = [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]
+# Reference pool: Used to generate rich dataset context (~50k tokens) in the meta-prompt
+# Must not overlap with bootstrap/online/test pools to avoid data leakage
+reference_pool = [50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149]
+
+# System specification (pipeline-level, covers both analyzer and classifier stages)
+spec_path = "examples/task_apps/banking77_pipeline/banking77_pipeline_spec.json"
+spec_max_tokens = 5000
+spec_include_examples = true
+spec_priority_threshold = 8  # Only include high-priority rules (8+)
+
+[prompt_learning.mipro.tpe]
+gamma = 0.25
+n_candidates = 32
+n_startup_trials = 10
+epsilon = 0.1
+categorical_alpha = 1.0
+prior_weight = 0.05
+
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 5
+sets_per_size = 6
+include_empty = true
+
+[prompt_learning.mipro.grounding]
+n = 10
+temperature = 0.7
+max_tokens = 600
+
+[prompt_learning.mipro.meta_update]
+enabled = true
+every_iterations = 3
+topk_success = 5
+topk_failure = 5
+validate_on_batch = 32
+keep_k = 12
+dedup_token_overlap = 0.8
+regen_temperature_decay = 0.95
+
+[[prompt_learning.mipro.modules]]
+module_id = "classifier"
+max_instruction_slots = 3
+max_demo_slots = 5
+
+[[prompt_learning.mipro.modules]]
+module_id = "calibrator"
+max_instruction_slots = 3
+max_demo_slots = 5

examples/blog_posts/mipro/configs/banking77_pipeline_mipro_test.toml

@@ -0,0 +1,170 @@
+# Lightweight smoke-test config for the Banking77 two-step pipeline
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "https://synth-laboratories-dev--synth-banking77-pipeline-web-web.modal.run"
+task_app_id = "banking77-pipeline"
+
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_pattern_test"
+name = "Banking77 Pipeline Pattern (Test)"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "Pipeline placeholder message; see metadata.pipeline_modules for concrete instructions."
+order = 0
+
+[prompt_learning.initial_prompt.metadata]
+pipeline_modules = [
+  {
+    name = "classifier",
+    instruction_text = "Classify the customer query into a Banking77 intent and invoke the tool.",
+    few_shots = [],
+  },
+  {
+    name = "calibrator",
+    instruction_text = "Validate or correct the classifier suggestion, then emit the final intent via the tool.",
+    few_shots = [],
+  },
+]
+
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "openai/gpt-oss-20b"
+provider = "groq"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-pipeline-test"
+
+[prompt_learning.env_config]
+pool = "train"
+
+[prompt_learning.mipro]
+env_name = "banking77_pipeline"
+num_iterations = 2
+num_evaluations_per_iteration = 2
+batch_size = 4
+max_concurrent = 8
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.75
+max_instructions = 2
+duplicate_retry_limit = 6
+bootstrap_train_seeds = [0, 1, 2, 3]
+online_pool = [4, 5, 6, 7]
+test_pool = [8, 9]
+
+[prompt_learning.mipro.tpe]
+gamma = 0.25
+n_candidates = 24
+n_startup_trials = 4
+epsilon = 0.2
+categorical_alpha = 1.0
+prior_weight = 0.05
+
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 3
+sets_per_size = 4
+include_empty = true
+
+[prompt_learning.mipro.grounding]
+n = 6
+temperature = 0.7
+max_tokens = 400
+
+[prompt_learning.mipro.meta_update]
+enabled = false
+# Banking77 Multi-Step Pipeline (MIPROv2) - Test Configuration
+# Reduced-iteration settings for CI and quick verification
+
+[prompt_learning]
+algorithm = "mipro"
+task_app_url = "http://127.0.0.1:8112"
+task_app_id = "banking77-pipeline"
+
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_pattern"
+name = "Banking77 Pipeline Classification Pattern"
+
+[[prompt_learning.initial_prompt.messages]]
+role = "system"
+pattern = "You orchestrate a two-stage Banking77 pipeline. The classifier proposes an intent and the calibrator verifies it. Always return the final intent label using the `banking77_classify` tool."
+order = 0
+
+[[prompt_learning.initial_prompt.messages]]
+role = "user"
+pattern = "Customer Query: {query}\n\nEnsure both classifier and calibrator stay within the Banking77 intent set."
+order = 1
+
+[prompt_learning.initial_prompt.wildcards]
+query = "REQUIRED"
+
+[prompt_learning.initial_prompt.metadata]
+pipeline_modules = [
+  { name = "classifier", instruction_text = "Map the customer query to the closest Banking77 intent. Stay precise, lean on the official label list, and respond via `banking77_classify`.", few_shots = [] },
+  { name = "calibrator", instruction_text = "Audit the classifier suggestion. Confirm when it matches the query; otherwise pick the best Banking77 intent and reply with `banking77_classify`.", few_shots = [] }
+]
+
+[prompt_learning.policy]
+inference_mode = "synth_hosted"
+model = "openai/gpt-oss-20b"
+provider = "groq"
+temperature = 0.0
+max_completion_tokens = 128
+policy_name = "banking77-pipeline-mipro"
+
+[prompt_learning.env_config]
+pool = "train"
+
+[prompt_learning.mipro]
+env_name = "banking77-pipeline"
+num_iterations = 4
+num_evaluations_per_iteration = 3
+batch_size = 3
+max_concurrent = 8
+meta_model = "gpt-4o-mini"
+meta_model_provider = "openai"
+few_shot_score_threshold = 0.80
+max_instructions = 3
+
+bootstrap_train_seeds = [0, 1, 2, 3]
+online_pool = [4, 5, 6, 7]
+test_pool = [8, 9, 10, 11]
+
+[prompt_learning.mipro.tpe]
+gamma = 0.25
+n_candidates = 16
+n_startup_trials = 5
+epsilon = 0.1
+categorical_alpha = 1.0
+prior_weight = 0.05
+
+[prompt_learning.mipro.demo]
+max_few_shot_examples = 4
+sets_per_size = 3
+include_empty = true
+
+[prompt_learning.mipro.grounding]
+n = 6
+temperature = 0.7
+max_tokens = 500
+
+[prompt_learning.mipro.meta_update]
+enabled = true
+every_iterations = 2
+topk_success = 3
+topk_failure = 3
+validate_on_batch = 12
+keep_k = 8
+dedup_token_overlap = 0.8
+regen_temperature_decay = 0.95
+
+[[prompt_learning.mipro.modules]]
+module_id = "classifier"
+max_instruction_slots = 2
+max_demo_slots = 4
+
+[[prompt_learning.mipro.modules]]
+module_id = "calibrator"
+max_instruction_slots = 2
+max_demo_slots = 4

examples/blog_posts/mipro/deploy_banking77_pipeline_task_app.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+# Deploy the Banking77 pipeline task app locally via uvicorn
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../../.." && pwd)"
+cd "$REPO_ROOT"
+
+PORT="8112"
+
+echo "🚀 Launching banking77-pipeline task app on port ${PORT}"
+uvx synth-ai deploy banking77-pipeline \
+  --runtime uvicorn \
+  --port "${PORT}" \
+  --env-file .env \
+  --follow
+#!/bin/bash
+# Deploy the Banking77 multi-step pipeline task app locally for MIPROv2 optimisation
+
+set -e
+
+echo "🚀 Deploying Banking77 Pipeline Task App for MIPROv2..."
+echo "======================================================"
+
+# Ensure ENVIRONMENT_API_KEY exists (shared with backend + CLI)
+export ENVIRONMENT_API_KEY="${ENVIRONMENT_API_KEY:-$(python -c 'import secrets; print(secrets.token_urlsafe(32))')}"
+
+# Optional providers (helpful during local experimentation)
+if [ -z "$GROQ_API_KEY" ] && [ -z "$OPENAI_API_KEY" ]; then
+    echo "⚠️  Neither GROQ_API_KEY nor OPENAI_API_KEY is set."
+    echo "   The task app can still start, but hosted inference may fail."
+else
+    if [ -n "$GROQ_API_KEY" ]; then
+        echo "✅ GROQ_API_KEY: ${GROQ_API_KEY:0:20}..."
+    fi
+    if [ -n "$OPENAI_API_KEY" ]; then
+        echo "✅ OPENAI_API_KEY: ${OPENAI_API_KEY:0:20}..."
+    fi
+fi
+
+echo "✅ ENVIRONMENT_API_KEY: ${ENVIRONMENT_API_KEY:0:20}..."
+
+# Navigate to repo root
+cd "$(dirname "$0")/../../.."
+
+echo ""
+echo "📦 Ensuring dependencies are installed..."
+uv pip install -e . --quiet || true
+
+echo ""
+echo "🏃 Starting Banking77 pipeline task app on http://127.0.0.1:8112"
+echo "Press Ctrl+C to stop."
+echo ""
+
+python -m examples.task_apps.banking77_pipeline.banking77_pipeline_task_app \
+    --host 0.0.0.0 \
+    --port 8112 \
+    --reload
+