ai-pipeline-core 0.1.13__py3-none-any.whl → 0.2.0__py3-none-any.whl

ai_pipeline_core/flow/config.py

@@ -10,11 +10,16 @@ Best Practice:
     to ensure type safety and proper validation of output documents.
 """
 
+import json
 from abc import ABC
 from typing import Any, ClassVar, Iterable
 
-from ai_pipeline_core.documents import DocumentList, FlowDocument
+from ai_pipeline_core.documents import Document, DocumentList, FlowDocument
 from ai_pipeline_core.exceptions import DocumentValidationError
+from ai_pipeline_core.logging import get_pipeline_logger
+from ai_pipeline_core.storage import Storage
+
+logger = get_pipeline_logger(__name__)
 
 
 class FlowConfig(ABC):
@@ -51,8 +56,10 @@ class FlowConfig(ABC):
         ...     OUTPUT_DOCUMENT_TYPE = ProcessedDocument  # Different type!
         >>>
         >>> # Use in @pipeline_flow - RECOMMENDED PATTERN
-        >>> @pipeline_flow(name="processing")
-        >>> async def process(config: ProcessingFlowConfig, docs: DocumentList) -> DocumentList:
+        >>> @pipeline_flow(config=ProcessingFlowConfig, name="processing")
+        >>> async def process(
+        ...     project_name: str, docs: DocumentList, flow_options: FlowOptions
+        ... ) -> DocumentList:
         ...     outputs = []
         ...     # ... processing logic ...
         ...     return config.create_and_validate_output(outputs)
@@ -289,8 +296,10 @@ class FlowConfig(ABC):
             DocumentValidationError: If output type doesn't match OUTPUT_DOCUMENT_TYPE.
 
         Example:
-            >>> @pipeline_flow(name="my_flow")
-            >>> async def process_flow(config: MyFlowConfig, ...) -> DocumentList:
+            >>> @pipeline_flow(config=MyFlowConfig, name="my_flow")
+            >>> async def process_flow(
+            ...     project_name: str, documents: DocumentList, flow_options: FlowOptions
+            ... ) -> DocumentList:
             >>>     outputs = []
             >>>     # ... processing logic ...
             >>>     outputs.append(OutputDoc(...))
@@ -312,3 +321,163 @@ class FlowConfig(ABC):
             documents = DocumentList(list(output))  # type: ignore[arg-type]
         cls.validate_output_documents(documents)
         return documents
+
+    @classmethod
+    async def load_documents(
+        cls,
+        uri: str,
+    ) -> DocumentList:
+        """Load documents from storage matching INPUT_DOCUMENT_TYPES.
+
+        Loads documents from a storage location based on the class's INPUT_DOCUMENT_TYPES.
+        Supports both local filesystem and Google Cloud Storage backends.
+        Automatically loads metadata (.description.md and .sources.json) when present.
+
+        Args:
+            uri: Storage URI (file://, gs://, or local path)
+
+        Returns:
+            DocumentList containing loaded documents matching INPUT_DOCUMENT_TYPES
+
+        Example:
+            >>> # Load from local filesystem
+            >>> docs = await MyFlowConfig.load_documents("./data")
+            >>>
+            >>> # Load from GCS (uses GCS_SERVICE_ACCOUNT_FILE from settings if configured)
+            >>> docs = await MyFlowConfig.load_documents("gs://bucket/data")
+        """
+        # Use INPUT_DOCUMENT_TYPES if not specified
+        storage = await Storage.from_uri(uri)
+        loaded_documents = DocumentList()
+
+        # Process each document type
+        for doc_type in cls.INPUT_DOCUMENT_TYPES:
+            canonical_name = doc_type.canonical_name()
+            doc_storage = storage.with_base(canonical_name)
+
+            # Check if subdirectory exists
+            if not await doc_storage.exists(""):
+                logger.debug(f"Subdirectory {canonical_name} not found, skipping")
+                continue
+
+            # List files in subdirectory
+            objects = await doc_storage.list("", recursive=False, include_dirs=False)
+
+            # Create lookup set for metadata files
+            object_keys = {obj.key for obj in objects}
+
+            # Filter out metadata files
+            doc_files = [
+                obj
+                for obj in objects
+                if not obj.key.endswith(Document.DESCRIPTION_EXTENSION)
+                and not obj.key.endswith(Document.SOURCES_EXTENSION)
+            ]
+
+            for obj in doc_files:
+                try:
+                    # Load document content
+                    content = await doc_storage.read_bytes(obj.key)
+
+                    # Load metadata if present
+                    description = None
+                    sources: list[str] = []
+
+                    # Check for description in objects list
+                    desc_path = f"{obj.key}{Document.DESCRIPTION_EXTENSION}"
+                    if desc_path in object_keys:
+                        try:
+                            description = await doc_storage.read_text(desc_path)
+                        except Exception as e:
+                            logger.warning(f"Failed to load description for {obj.key}: {e}")
+
+                    # Check for sources in objects list
+                    sources_path = f"{obj.key}{Document.SOURCES_EXTENSION}"
+                    if sources_path in object_keys:
+                        try:
+                            sources_text = await doc_storage.read_text(sources_path)
+                            sources = json.loads(sources_text)
+                        except Exception as e:
+                            logger.warning(f"Failed to load sources for {obj.key}: {e}")
+
+                    # Create document instance
+                    doc = doc_type(
+                        name=obj.key,
+                        content=content,
+                        description=description,
+                        sources=sources,
+                    )
+
+                    loaded_documents.append(doc)
+                    logger.debug(f"Loaded {doc_type.__name__} document: {obj.key}")
+                except Exception as e:
+                    logger.error(f"Failed to load {doc_type.__name__} document {obj.key}: {e}")
+
+        logger.info(f"Loaded {len(loaded_documents)} documents from {uri}")
+        return loaded_documents
+
+    @classmethod
+    async def save_documents(
+        cls,
+        uri: str,
+        documents: DocumentList,
+        *,
+        validate_output_type: bool = True,
+    ) -> None:
+        """Save documents to storage with metadata.
+
+        Saves FlowDocument instances to a storage location with their content
+        and metadata files (Document.DESCRIPTION_EXTENSION and Document.SOURCES_EXTENSION).
+        Non-FlowDocument instances (TaskDocument, TemporaryDocument) are skipped.
+
+        Args:
+            uri: Storage URI (file://, gs://, or local path)
+            documents: DocumentList to save
+            validate_output_type: If True, validate documents match cls.OUTPUT_DOCUMENT_TYPE
+
+        Raises:
+            DocumentValidationError: If validate_output_type=True and documents don't match
+                                   OUTPUT_DOCUMENT_TYPE
+
+        Example:
+            >>> # Save to local filesystem
+            >>> await MyFlowConfig.save_documents("./output", docs)
+            >>>
+            >>> # Save to GCS (uses GCS_SERVICE_ACCOUNT_FILE from settings if configured)
+            >>> await MyFlowConfig.save_documents("gs://bucket/output", docs)
+        """
+        # Validate output type if requested
+        if validate_output_type:
+            cls.validate_output_documents(documents)
+
+        storage = await Storage.from_uri(uri)
+        saved_count = 0
+
+        for doc in documents:
+            # Skip non-FlowDocument instances
+            if not isinstance(doc, FlowDocument):
+                logger.warning(f"Skipping non-FlowDocument: {type(doc).__name__}")
+                continue
+
+            # Get canonical name for subdirectory
+            canonical_name = doc.canonical_name()
+            doc_storage = storage.with_base(canonical_name)
+
+            # Save document content
+            await doc_storage.write_bytes(doc.name, doc.content)
+            saved_count += 1
+
+            # Save description if present
+            if doc.description:
+                desc_path = f"{doc.name}{Document.DESCRIPTION_EXTENSION}"
+                await doc_storage.write_text(desc_path, doc.description)
+
+            # Save sources if present
+            if doc.sources:
+                sources_path = f"{doc.name}{Document.SOURCES_EXTENSION}"
+                sources_json = json.dumps(doc.sources, indent=2)
+                await doc_storage.write_text(sources_path, sources_json)
+
+            logger.debug(f"Saved {type(doc).__name__} document: {doc.name}")
+
+        logger.info(f"Saved {saved_count} documents to {uri}")

ai_pipeline_core/flow/options.py

@@ -60,11 +60,11 @@ class FlowOptions(BaseSettings):
         add flow-specific parameters with appropriate validation.
     """
 
-    core_model: ModelName | str = Field(
+    core_model: ModelName = Field(
         default="gpt-5",
         description="Primary model for complex analysis and generation tasks.",
     )
-    small_model: ModelName | str = Field(
+    small_model: ModelName = Field(
         default="gpt-5-mini",
         description="Fast, cost-effective model for simple tasks and orchestration.",
     )

ai_pipeline_core/llm/__init__.py

@@ -8,6 +8,8 @@ from .ai_messages import AIMessages, AIMessageType
 from .client import (
     generate,
     generate_structured,
+    generate_with_retry_for_testing,
+    process_messages_for_testing,
 )
 from .model_options import ModelOptions
 from .model_response import ModelResponse, StructuredModelResponse
@@ -17,9 +19,12 @@ __all__ = [
     "AIMessages",
     "AIMessageType",
     "ModelName",
-    "ModelOptions",
     "ModelResponse",
+    "ModelOptions",
     "StructuredModelResponse",
     "generate",
     "generate_structured",
+    # Internal functions exposed for testing only
+    "process_messages_for_testing",
+    "generate_with_retry_for_testing",
 ]

ai_pipeline_core/llm/ai_messages.py

@@ -48,22 +48,31 @@ class AIMessages(list[AIMessageType]):
         - ModelResponse: Becomes {"role": "assistant", "content": response.content}
 
     Note: Document conversion is automatic. Text content becomes user text messages.
-    Images are sent to vision-capable models (non-vision models will raise ValueError).
-    PDFs are attached when supported by the model, otherwise a text extraction
-    fallback is used. LiteLLM proxy handles the specific encoding requirements
-    for each provider.
+
+    VISION/PDF MODEL COMPATIBILITY WARNING:
+    Images require vision-capable models (e.g., gpt-4o, gemini-pro-vision, claude-3-haiku).
+    Non-vision models will raise ValueError when encountering image documents.
+    PDFs require models with document processing support - check your model's capabilities
+    before including PDF documents in messages. Unsupported models may fall back to
+    text extraction or raise errors depending on provider configuration.
+    LiteLLM proxy handles the specific encoding requirements for each provider.
 
     IMPORTANT: Although AIMessages can contain Document entries, the LLM client functions
     expect `messages` to be `AIMessages` or `str`. If you start from a Document or a list
     of Documents, build AIMessages first (e.g., `AIMessages([doc])` or `AIMessages(docs)`).
 
+    CAUTION: AIMessages is a list subclass. Always use list construction (e.g.,
+    `AIMessages(["text"])`) or empty constructor with append (e.g.,
+    `AIMessages(); messages.append("text")`). Never pass raw strings directly to the
+    constructor (`AIMessages("text")`) as this will iterate over the string characters
+    instead of treating it as a single message.
+
     Example:
         >>> from ai_pipeline_core import llm
         >>> messages = AIMessages()
         >>> messages.append("What is the capital of France?")
         >>> response = await llm.generate("gpt-5", messages=messages)
         >>> messages.append(response)  # Add the actual response
-        >>> prompt = messages.get_last_message_as_str()  # Get the last message as a string
     """
 
     def get_last_message(self) -> AIMessageType:
@@ -78,8 +87,6 @@ class AIMessages(list[AIMessageType]):
     def get_last_message_as_str(self) -> str:
         """Get the last message as a string, raising if not a string.
 
-        @public
-
         Returns:
             The last message as a string.
 

ai_pipeline_core/llm/client.py

@@ -24,7 +24,6 @@ from pydantic import BaseModel
 
 from ai_pipeline_core.exceptions import LLMError
 from ai_pipeline_core.settings import settings
-from ai_pipeline_core.tracing import trace
 
 from .ai_messages import AIMessages
 from .model_options import ModelOptions
@@ -38,6 +37,7 @@ def _process_messages(
     context: AIMessages,
     messages: AIMessages,
     system_prompt: str | None = None,
+    cache_ttl: str | None = "120s",
 ) -> list[ChatCompletionMessageParam]:
     """Process and format messages for LLM API consumption.
 
@@ -49,21 +49,25 @@ def _process_messages(
         context: Messages to be cached (typically expensive/static content).
         messages: Regular messages without caching (dynamic queries).
         system_prompt: Optional system instructions for the model.
+        cache_ttl: Cache TTL for context messages (e.g. "120s", "5m", "1h").
+                   Set to None or empty string to disable caching.
 
     Returns:
         List of formatted messages ready for API calls, with:
         - System prompt at the beginning (if provided)
-        - Context messages with cache_control on the last one
+        - Context messages with cache_control on the last one (if cache_ttl)
         - Regular messages without caching
 
     System Prompt Location:
-        The system prompt from ModelOptions.system_prompt is always injected
-        as the FIRST message with role="system". It is NOT cached with context,
-        allowing dynamic system prompts without breaking cache efficiency.
+        The system prompt parameter is always injected as the FIRST message
+        with role="system". It is NOT cached with context, allowing dynamic
+        system prompts without breaking cache efficiency.
 
     Cache behavior:
-        The last context message gets ephemeral caching (120s TTL)
+        The last context message gets ephemeral caching with specified TTL
         to reduce token usage on repeated calls with same context.
+        If cache_ttl is None or empty string (falsy), no caching is applied.
+        Only the last context message receives cache_control to maximize efficiency.
 
     Note:
         This is an internal function used by _generate_with_retry().
@@ -80,11 +84,12 @@ def _process_messages(
         # Use AIMessages.to_prompt() for context
         context_messages = context.to_prompt()
 
-        # Apply caching to last context message
-        context_messages[-1]["cache_control"] = {  # type: ignore
-            "type": "ephemeral",
-            "ttl": "120s",  # Cache for 2m
-        }
+        # Apply caching to last context message if cache_ttl is set
+        if cache_ttl:
+            context_messages[-1]["cache_control"] = {  # type: ignore
+                "type": "ephemeral",
+                "ttl": cache_ttl,
+            }
 
         processed_messages.extend(context_messages)
 
@@ -173,7 +178,9 @@ async def _generate_with_retry(
     if not context and not messages:
         raise ValueError("Either context or messages must be provided")
 
-    processed_messages = _process_messages(context, messages, options.system_prompt)
+    processed_messages = _process_messages(
+        context, messages, options.system_prompt, options.cache_ttl
+    )
     completion_kwargs: dict[str, Any] = {
         "model": model,
         "messages": processed_messages,
@@ -213,9 +220,8 @@ async def _generate_with_retry(
     raise LLMError("Unknown error occurred during LLM generation.")
 
 
-@trace(ignore_inputs=["context"])
 async def generate(
-    model: ModelName | str,
+    model: ModelName,
     *,
     context: AIMessages | None = None,
     messages: AIMessages | str,
@@ -230,20 +236,24 @@ async def generate(
     expensive static content separately from dynamic queries.
 
     Best Practices:
-        1. OPTIONS: Omit in 90% of cases - defaults are optimized
+        1. OPTIONS: DO NOT use the options parameter - omit it entirely for production use
         2. MESSAGES: Use AIMessages or str - wrap Documents in AIMessages
         3. CONTEXT vs MESSAGES: Use context for static/cacheable, messages for dynamic
+        4. CONFIGURATION: Configure model behavior via LiteLLM proxy or environment variables
 
     Args:
         model: Model to use (e.g., "gpt-5", "gemini-2.5-pro", "grok-4").
-               Can be ModelName literal or any string for custom models.
+               Accepts predefined models or any string for custom models.
         context: Static context to cache (documents, examples, instructions).
                 Defaults to None (empty context). Cached for 120 seconds.
         messages: Dynamic messages/queries. AIMessages or str ONLY.
                  Do not pass Document or DocumentList directly.
                  If string, converted to AIMessages internally.
-        options: Model configuration (temperature, retries, timeout, etc.).
-                Defaults to None (uses ModelOptions() with standard settings).
+        options: DEPRECATED - DO NOT USE. Reserved for internal framework usage only.
+                Framework defaults are production-optimized (3 retries, 10s delay, 300s timeout).
+                Configure model behavior centrally via LiteLLM proxy settings or environment
+                variables, not per API call. Provider-specific settings should be configured
+                at the proxy level.
 
     Returns:
         ModelResponse containing:
@@ -268,17 +278,26 @@ async def generate(
         # WRONG - don't convert to string yourself
         response = await llm.generate("gpt-5", messages=my_document.text)  # NO!
 
+    VISION/PDF MODEL COMPATIBILITY:
+        When using Documents containing images or PDFs, ensure your model supports these formats:
+        - Images require vision-capable models (gpt-4o, gemini-pro-vision, claude-3-sonnet)
+        - PDFs require document processing support (varies by provider)
+        - Non-compatible models will raise ValueError or fall back to text extraction
+        - Check model capabilities before including visual/PDF content
+
     Context vs Messages Strategy:
-        context: Static, reusable content (cached 120 seconds)
+        context: Static, reusable content for caching efficiency
             - Large documents, instructions, examples
-            - Same across multiple calls
+            - Remains constant across multiple calls
+            - Cached when supported by provider/proxy configuration
 
-        messages: Dynamic, query-specific content
+        messages: Dynamic, per-call specific content
             - User questions, current conversation turn
-            - Changes every call
+            - Changes with each API call
+            - Never cached, always processed fresh
 
     Example:
-        >>> # Simple case - no options needed (90% of cases)
+        >>> # CORRECT - No options parameter (this is the recommended pattern)
         >>> response = await llm.generate("gpt-5", messages="Explain quantum computing")
         >>> print(response.content)  # In production, use get_pipeline_logger instead of print
 
@@ -292,13 +311,6 @@ async def generate(
         >>> # Second call: reuses cache, saves tokens!
         >>> r2 = await llm.generate("gpt-5", context=static_doc, messages="Key points?")
 
-        >>> # AVOID unnecessary options (defaults are optimal)
-        >>> response = await llm.generate(
-        ...     "gpt-5",
-        ...     messages="Hello",
-        ...     options=ModelOptions(temperature=0.7)  # Default is probably fine!
-        ... )
-
         >>> # Multi-turn conversation
         >>> messages = AIMessages([
         ...     "What is Python?",
@@ -307,28 +319,48 @@ async def generate(
         ... ])
         >>> response = await llm.generate("gpt-5", messages=messages)
 
+        Configuration via LiteLLM Proxy:
+        >>> # Configure temperature in litellm_config.yaml:
+        >>> # model_list:
+        >>> #   - model_name: gpt-5
+        >>> #     litellm_params:
+        >>> #       model: openai/gpt-4o
+        >>> #       temperature: 0.3
+        >>> #       max_tokens: 1000
+        >>>
+        >>> # Configure retry logic in proxy:
+        >>> # general_settings:
+        >>> #   master_key: sk-1234
+        >>> #   max_retries: 5
+        >>> #   retry_delay: 15
+
     Performance:
         - Context caching saves ~50-90% tokens on repeated calls
         - First call: full token cost
-        - Subsequent calls (within 120s): only messages tokens
-        - Default retry delay is 10s (configurable via ModelOptions.retry_delay_seconds)
+        - Subsequent calls (within cache TTL): only messages tokens
+        - Default cache TTL is 120s (production-optimized)
+        - Default retry logic: 3 attempts with 10s delay (production-optimized)
 
     Caching:
         When enabled in your LiteLLM proxy and supported by the upstream provider,
-        context messages may be cached (typical TTL ~120s) to reduce token usage on
-        repeated calls. Savings depend on provider and payload; treat this as an
-        optimization, not a guarantee. Cache behavior varies by proxy configuration.
+        context messages may be cached to reduce token usage on repeated calls.
+        Default TTL is 120s (optimized for production workloads). Configure caching
+        behavior centrally via your LiteLLM proxy settings, not per API call.
+        Savings depend on provider and payload; treat this as an optimization, not a guarantee.
+
+    Configuration:
+        All model behavior should be configured at the LiteLLM proxy level:
+        - Temperature, max_tokens: Set in litellm_config.yaml model_list
+        - Retry logic: Configure in proxy general_settings
+        - Timeouts: Set via proxy configuration
+        - Caching: Enable/configure in proxy cache settings
+
+        This centralizes configuration and ensures consistency across all API calls.
 
     Note:
-        - Context argument is ignored by the tracer to avoid recording large data
         - All models are accessed via LiteLLM proxy
         - Automatic retry with configurable delay between attempts
         - Cost tracking via response headers
-
-    See Also:
-        - generate_structured: For typed/structured output
-        - AIMessages: Message container with document support
-        - ModelOptions: Configuration options
     """
     if isinstance(messages, str):
         messages = AIMessages([messages])
@@ -348,9 +380,8 @@ T = TypeVar("T", bound=BaseModel)
 """Type variable for Pydantic model types in structured generation."""
 
 
-@trace(ignore_inputs=["context"])
 async def generate_structured(
-    model: ModelName | str,
+    model: ModelName,
     response_format: type[T],
     *,
     context: AIMessages | None = None,
@@ -364,20 +395,71 @@ async def generate_structured(
     Type-safe generation that returns validated Pydantic model instances.
     Uses OpenAI's structured output feature for guaranteed schema compliance.
 
-    Best Practices (same as generate):
-        1. OPTIONS: Omit in 90% of cases - defaults are optimized
+    IMPORTANT: Search models (models with '-search' suffix) do not support
+    structured output. Use generate() instead for search models.
+
+    Best Practices:
+        1. OPTIONS: DO NOT use the options parameter - omit it entirely for production use
         2. MESSAGES: Use AIMessages or str - wrap Documents in AIMessages
-        3. CONTEXT vs MESSAGES: Use context for static/cacheable, messages for dynamic
+        3. CONFIGURATION: Configure model behavior via LiteLLM proxy or environment variables
+        4. See generate() documentation for more details
+
+    Context vs Messages Strategy:
+        context: Static, reusable content for caching efficiency
+            - Schemas, examples, instructions
+            - Remains constant across multiple calls
+            - Cached when supported by provider/proxy configuration
+
+        messages: Dynamic, per-call specific content
+            - Data to be structured, user queries
+            - Changes with each API call
+            - Never cached, always processed fresh
+
+    Complex Task Pattern:
+        For complex tasks like research or deep analysis, it's recommended to use
+        a two-step approach:
+        1. First use generate() with a capable model to perform the analysis
+        2. Then use generate_structured() with a smaller model to convert the
+           response into structured output
+
+        This pattern is more reliable than trying to force complex reasoning
+        directly into structured format:
+
+        >>> # Step 1: Research/analysis with generate() - no options parameter
+        >>> research = await llm.generate(
+        ...     "gpt-5",
+        ...     messages="Research and analyze this complex topic..."
+        ... )
+        >>>
+        >>> # Step 2: Structure the results with generate_structured()
+        >>> structured = await llm.generate_structured(
+        ...     "gpt-5-mini",  # Smaller model is fine for structuring
+        ...     response_format=ResearchSummary,
+        ...     messages=f"Extract key information: {research.content}"
+        ... )
 
     Args:
         model: Model to use (must support structured output).
+               Search models (models with '-search' suffix) do not support structured output.
         response_format: Pydantic model class defining the output schema.
                         The model will generate JSON matching this schema.
         context: Static context to cache (documents, schemas, examples).
                 Defaults to None (empty AIMessages).
         messages: Dynamic prompts/queries. AIMessages or str ONLY.
                  Do not pass Document or DocumentList directly.
-        options: Model configuration. response_format is set automatically.
+        options: DEPRECATED - DO NOT USE. Reserved for internal framework usage only.
+                Framework defaults are production-optimized. Configure model behavior
+                centrally via LiteLLM proxy settings, not per API call.
+                The response_format is set automatically from the response_format parameter.
+
+    VISION/PDF MODEL COMPATIBILITY:
+        When using Documents with images/PDFs in structured output:
+        - Images require vision-capable models that also support structured output
+        - PDFs require models with both document processing AND structured output support
+        - Many models support either vision OR structured output, but not both
+        - Test your specific model+document combination before production use
+        - Consider two-step approach: generate() for analysis, then generate_structured()
+          for formatting
 
     Returns:
         StructuredModelResponse[T] containing:
@@ -387,6 +469,7 @@ async def generate_structured(
     Raises:
         TypeError: If response_format is not a Pydantic model class.
         ValueError: If model doesn't support structured output or no parsed content returned.
+                   Structured output support varies by provider and model.
         LLMError: If generation fails after retries.
         ValidationError: If response cannot be parsed into response_format.
 
@@ -398,8 +481,9 @@ async def generate_structured(
         ...     sentiment: float = Field(ge=-1, le=1)
         ...     key_points: list[str] = Field(max_length=5)
         >>>
+        >>> # CORRECT - No options parameter
         >>> response = await llm.generate_structured(
-        ...     model="gpt-5",
+        ...     "gpt-5",
         ...     response_format=Analysis,
         ...     messages="Analyze this product review: ..."
         ... )
@@ -410,11 +494,13 @@ async def generate_structured(
         ...     print(f"- {point}")
 
     Supported models:
-        Support varies by provider and model. Generally includes:
+        Structured output support varies by provider and model. Generally includes:
         - OpenAI: GPT-4 and newer models
         - Anthropic: Claude 3+ models
         - Google: Gemini Pro models
-        Check provider documentation for specific model support.
+
+        Search models (models with '-search' suffix) do not support structured output.
+        Check provider documentation for specific support.
 
     Performance:
         - Structured output may use more tokens than free text
@@ -426,11 +512,7 @@ async def generate_structured(
         - The model generates JSON matching the schema
         - Validation happens automatically via Pydantic
         - Use Field() descriptions to guide generation
-
-    See Also:
-        - generate: For unstructured text generation
-        - ModelOptions: Configuration including response_format
-        - StructuredModelResponse: Response wrapper with .parsed property
+        - Search models (models with '-search' suffix) do not support structured output
     """
     if context is None:
         context = AIMessages()
@@ -473,3 +555,9 @@ async def generate_structured(
 
     # Create a StructuredModelResponse with the parsed value
     return StructuredModelResponse[T](chat_completion=response, parsed_value=parsed_value)
+
+
+# Public aliases for testing internal functions
+# These are exported to allow testing of implementation details
+process_messages_for_testing = _process_messages
+generate_with_retry_for_testing = _generate_with_retry