From 52adedab4affcd4b576fa78f29972f8d2b989c7e Mon Sep 17 00:00:00 2001 From: ValueOn AG Date: Tue, 21 Oct 2025 17:12:13 +0200 Subject: [PATCH] architecture for looping ai calls tested and fixed --- modules/connectors/connectorAiAnthropic.py | 25 +++++++++-- modules/connectors/connectorAiOpenai.py | 6 ++- modules/connectors/connectorAiPerplexity.py | 10 ++--- modules/interfaces/interfaceAiObjects.py | 42 ++++++++++++------- modules/services/serviceAi/subCoreAi.py | 29 ++++++++++--- .../mainServiceGeneration.py | 31 +++++++++++--- .../serviceGeneration/subPromptBuilder.py | 10 ++--- test_ai_behavior.py | 8 +--- 8 files changed, 113 insertions(+), 48 deletions(-) diff --git a/modules/connectors/connectorAiAnthropic.py b/modules/connectors/connectorAiAnthropic.py index 85cf62f2..16b235e0 100644 --- a/modules/connectors/connectorAiAnthropic.py +++ b/modules/connectors/connectorAiAnthropic.py @@ -39,6 +39,26 @@ class AiAnthropic: logger.info(f"Anthropic Connector initialized with model: {self.modelName}") + def _getMaxTokensForModel(self, maxTokens: int = None) -> int: + """Get appropriate max_tokens for the current model.""" + if maxTokens is not None: + return maxTokens + + # Model-specific defaults based on Anthropic's limits + model_name = self.modelName.lower() + if "claude-3-5-sonnet" in model_name: + return 200000 # Claude 3.5 Sonnet max + elif "claude-3-5-haiku" in model_name: + return 200000 # Claude 3.5 Haiku max + elif "claude-3-opus" in model_name: + return 200000 # Claude 3 Opus max + elif "claude-3-sonnet" in model_name: + return 200000 # Claude 3 Sonnet max + elif "claude-3-haiku" in model_name: + return 200000 # Claude 3 Haiku max + else: + return 200000 # Default to maximum for unknown models + async def callAiBasic(self, messages: List[Dict[str, Any]], temperature: float = None, maxTokens: int = None) -> Dict[str, Any]: """ Calls the Anthropic API with the given messages. @@ -106,9 +126,8 @@ class AiAnthropic: "temperature": temperature, } - # Only add max_tokens if it's explicitly set - if maxTokens is not None: - payload["max_tokens"] = maxTokens + # Anthropic requires max_tokens - use model-appropriate value + payload["max_tokens"] = self._getMaxTokensForModel(maxTokens) if system_prompt: payload["system"] = system_prompt diff --git a/modules/connectors/connectorAiOpenai.py b/modules/connectors/connectorAiOpenai.py index c768888e..8aac34cd 100644 --- a/modules/connectors/connectorAiOpenai.py +++ b/modules/connectors/connectorAiOpenai.py @@ -80,7 +80,8 @@ class AiOpenai: ) if response.status_code != 200: - logger.error(f"OpenAI API error: {response.status_code} - {response.text}") + error_message = f"OpenAI API error: {response.status_code} - {response.text}" + logger.error(error_message) # Check for context length exceeded error if response.status_code == 400: @@ -95,7 +96,8 @@ class AiOpenai: except (ValueError, KeyError): pass # If we can't parse the error, fall through to generic error - raise HTTPException(status_code=500, detail="Error communicating with OpenAI API") + # Include the actual error details in the exception + raise HTTPException(status_code=500, detail=error_message) responseJson = response.json() content = responseJson["choices"][0]["message"]["content"] diff --git a/modules/connectors/connectorAiPerplexity.py b/modules/connectors/connectorAiPerplexity.py index b075a84d..4634cf1d 100644 --- a/modules/connectors/connectorAiPerplexity.py +++ b/modules/connectors/connectorAiPerplexity.py @@ -68,9 +68,8 @@ class AiPerplexity: "temperature": temperature } - # Only add max_tokens if it's explicitly set - if maxTokens is not None: - payload["max_tokens"] = maxTokens + # Add max_tokens - use provided value or default to 128000 (Perplexity's typical limit) + payload["max_tokens"] = maxTokens if maxTokens is not None else 128000 response = await self.httpClient.post( self.apiUrl, @@ -135,9 +134,8 @@ class AiPerplexity: "temperature": temperature } - # Only add max_tokens if it's explicitly set - if maxTokens is not None: - payload["max_tokens"] = maxTokens + # Add max_tokens - use provided value or default to 128000 (Perplexity's typical limit) + payload["max_tokens"] = maxTokens if maxTokens is not None else 128000 response = await self.httpClient.post( self.apiUrl, diff --git a/modules/interfaces/interfaceAiObjects.py b/modules/interfaces/interfaceAiObjects.py index 2e9b6b38..337a2878 100644 --- a/modules/interfaces/interfaceAiObjects.py +++ b/modules/interfaces/interfaceAiObjects.py @@ -525,25 +525,24 @@ class AiObjects: try: logger.info(f"Attempting AI call with model: {modelName} (attempt {attempt + 1}/{len(fallbackModels)})") - # Store the selected model for token limit resolution - self._lastSelectedModel = modelName - - # Replace placeholder in prompt and context if present + # Replace placeholder in prompt for this specific model context_length = aiModels[modelName].get("contextLength", 0) if context_length > 0: token_limit = str(context_length) else: - token_limit = "4000" # Default for text generation + token_limit = "16000" # Default for text generation - if "" in prompt: - prompt = prompt.replace("", token_limit) + # Create a copy of the prompt for this model call + modelPrompt = prompt + if "" in modelPrompt: + modelPrompt = modelPrompt.replace("", token_limit) logger.debug(f"Replaced with {token_limit} for model {modelName}") # Update messages array with replaced content messages = [] if context: messages.append({"role": "system", "content": f"Context from documents:\n{context}"}) - messages.append({"role": "user", "content": prompt}) + messages.append({"role": "user", "content": modelPrompt}) # Start timing startTime = time.time() @@ -562,19 +561,19 @@ class AiObjects: content = response["choices"][0]["message"]["content"] elif functionName == "callAiWithWebSearch": # Perplexity web search function - query = prompt + query = modelPrompt if context: - query = f"Context: {context}\n\nQuery: {prompt}" + query = f"Context: {context}\n\nQuery: {modelPrompt}" content = await connector.callAiWithWebSearch(query) elif functionName == "researchTopic": # Perplexity research function - content = await connector.researchTopic(prompt) + content = await connector.researchTopic(modelPrompt) elif functionName == "answerQuestion": # Perplexity question answering function - content = await connector.answerQuestion(prompt, context) + content = await connector.answerQuestion(modelPrompt, context) elif functionName == "getCurrentNews": # Perplexity news function - content = await connector.getCurrentNews(prompt) + content = await connector.getCurrentNews(modelPrompt) else: raise ValueError(f"Function {functionName} not supported for text generation") @@ -599,7 +598,14 @@ class AiObjects: except Exception as e: lastError = e - logger.warning(f"❌ AI call failed with model {modelName}: {str(e)}") + # Enhanced error logging with more details + error_details = str(e) + if hasattr(e, 'detail'): + error_details = f"{error_details} (detail: {e.detail})" + if hasattr(e, 'status_code'): + error_details = f"{error_details} (status: {e.status_code})" + + logger.warning(f"❌ AI call failed with model {modelName}: {error_details}") # If this is not the last model, try the next one if attempt < len(fallbackModels) - 1: @@ -611,7 +617,13 @@ class AiObjects: break # All fallback attempts failed - return error response - errorMsg = f"All AI models failed for operation {options.operationType}. Last error: {str(lastError)}" + last_error_details = str(lastError) + if hasattr(lastError, 'detail'): + last_error_details = f"{last_error_details} (detail: {lastError.detail})" + if hasattr(lastError, 'status_code'): + last_error_details = f"{last_error_details} (status: {lastError.status_code})" + + errorMsg = f"All AI models failed for operation {options.operationType}. Last error: {last_error_details}" logger.error(errorMsg) return AiCallResponse( content=errorMsg, diff --git a/modules/services/serviceAi/subCoreAi.py b/modules/services/serviceAi/subCoreAi.py index 6afaafce..07caa972 100644 --- a/modules/services/serviceAi/subCoreAi.py +++ b/modules/services/serviceAi/subCoreAi.py @@ -15,14 +15,14 @@ CRITICAL LIMITS: tokens total (reserve 20% for JSON structure) MANDATORY RULES: 1. STOP at approximately 80% of limit to ensure valid JSON completion 2. Return ONLY raw JSON (no ```json blocks, no text before/after) -3. ALWAYS include "continuation" field - this is MANDATORY CONTINUATION REQUIREMENTS: +Refer to the json object below where to set the "continuation" information: - If you can complete the full request: {"continuation": null} - If you must stop early: { "continuation": { - "last_data_items": "exact last items you generated (copy them exactly)", - "next_instruction": "Continue from [exact last item] - generate next items" + "last_data_items": "delivered last data for context (copy them)", + "next_instruction": "instruction for next data to deliver" } } @@ -101,7 +101,7 @@ class SubCoreAi: elif loopInstruction and iteration > 1: continuationContent = self._buildContinuationContent(accumulatedContent, iteration) if "LOOP_INSTRUCTION" in prompt: - iterationPrompt = prompt.replace("LOOP_INSTRUCTION", f"{loopInstruction}\n\n{continuationContent}") + iterationPrompt = prompt.replace("LOOP_INSTRUCTION", f"{continuationContent}\n\n{loopInstruction}") else: iterationPrompt = prompt else: @@ -208,10 +208,29 @@ class SubCoreAi: except (json.JSONDecodeError, KeyError, ValueError): pass + # Extract specific attributes from continuation object + last_data_items = "" + next_instruction = "" + + if continuation_description: + try: + if isinstance(continuation_description, str): + continuation_obj = json.loads(continuation_description) + else: + continuation_obj = continuation_description + + if isinstance(continuation_obj, dict): + last_data_items = continuation_obj.get("last_data_items", "") + next_instruction = continuation_obj.get("next_instruction", "") + except (json.JSONDecodeError, TypeError): + pass + continuation_content = f"""CONTINUATION REQUEST (Iteration {iteration}): You are continuing a previous response. DO NOT repeat any previous content. -{f"CONTINUATION INSTRUCTIONS: {continuation_description}" if continuation_description else "No specific continuation instructions provided."} +{f"Already delivered data: {last_data_items}" if last_data_items else "No previous data specified"} + +{f"Your task to deliver: {next_instruction}" if next_instruction else "No specific task provided"} CRITICAL REQUIREMENTS: - Start from the exact point specified in continuation instructions diff --git a/modules/services/serviceGeneration/mainServiceGeneration.py b/modules/services/serviceGeneration/mainServiceGeneration.py index cbb5cb0f..5d4489ac 100644 --- a/modules/services/serviceGeneration/mainServiceGeneration.py +++ b/modules/services/serviceGeneration/mainServiceGeneration.py @@ -315,10 +315,31 @@ class GenerationService: if not isinstance(extractedContent, dict): raise ValueError("extractedContent must be a JSON dictionary") - if "sections" not in extractedContent: - raise ValueError("extractedContent must contain 'sections' field") - - # Remove extra debug file writes for render inputs per simplification + # Check if this is a multi-document structure + if "documents" in extractedContent and len(extractedContent["documents"]) > 1: + # Multiple documents - use multi-file renderer + generated_documents = await self._renderMultiFileReport(extractedContent, outputFormat, title, userPrompt, aiService) + # For multi-document, return the first document's content and mime type + if generated_documents: + return generated_documents[0]["content"], generated_documents[0]["mime_type"] + else: + raise ValueError("No documents could be rendered") + elif "documents" in extractedContent and len(extractedContent["documents"]) == 1: + # Single document in documents array - extract sections + single_doc = extractedContent["documents"][0] + if "sections" not in single_doc: + raise ValueError("Document must contain 'sections' field") + # Create content for single document renderer + contentToRender = { + "sections": single_doc["sections"], + "metadata": extractedContent.get("metadata", {}), + "continuation": extractedContent.get("continuation", None) + } + elif "sections" in extractedContent: + # Direct sections format + contentToRender = extractedContent + else: + raise ValueError("extractedContent must contain 'sections' field or 'documents' array") # Get the appropriate renderer for the format renderer = self._getFormatRenderer(outputFormat) @@ -326,7 +347,7 @@ class GenerationService: raise ValueError(f"Unsupported output format: {outputFormat}") # Render the JSON content directly (AI generation handled by main service) - renderedContent, mimeType = await renderer.render(extractedContent, title, userPrompt, aiService) + renderedContent, mimeType = await renderer.render(contentToRender, title, userPrompt, aiService) # Remove extra debug output file writes logger.info(f"Successfully rendered JSON report to {outputFormat} format: {len(renderedContent)} characters") diff --git a/modules/services/serviceGeneration/subPromptBuilder.py b/modules/services/serviceGeneration/subPromptBuilder.py index 8f4afdb4..d326772c 100644 --- a/modules/services/serviceGeneration/subPromptBuilder.py +++ b/modules/services/serviceGeneration/subPromptBuilder.py @@ -19,7 +19,6 @@ logger = logging.getLogger(__name__) # Centralized JSON structure template for document generation JSON_STRUCTURE_TEMPLATE = """{ - "continuation": null, "metadata": { "title": "{{DOCUMENT_TITLE}}", "splitStrategy": "single_document", @@ -44,7 +43,8 @@ JSON_STRUCTURE_TEMPLATE = """{ "order": 1 } ] - }] + }], + "continuation": null, }""" async def buildAdaptiveExtractionPrompt( @@ -185,17 +185,17 @@ async def buildGenerationPrompt( # Always use the proper generation prompt template with LOOP_INSTRUCTION result = f"""Generate structured JSON content for document creation. -USER REQUEST: "{userPrompt}" +USER CONTEXT: "{userPrompt}" DOCUMENT TITLE: "{title}" TARGET FORMAT: {outputFormat} +LOOP_INSTRUCTION + RULES: - Follow the template structure below exactly; emit only one JSON object in the response - Fill sections with content based on the user request - Use appropriate content_type -LOOP_INSTRUCTION - Return ONLY valid JSON matching this structure (template below). Do not include any prose before/after. Use this as the single template reference for your output: {json_template} """ diff --git a/test_ai_behavior.py b/test_ai_behavior.py index cd6d1b32..b5951d06 100644 --- a/test_ai_behavior.py +++ b/test_ai_behavior.py @@ -334,13 +334,7 @@ class AIBehaviorTester: PROMPT_SCENARIOS = { "Prime Numbers Test": """Generate the first 5000 prime numbers in a table with 10 columns per row.""", - "Fibonacci Sequence": """Generate the first 1000 Fibonacci numbers in a table with 5 columns per row.""", - - "Multiplication Table": """Generate multiplication tables from 1 to 50, each table with 10 columns per row.""", - - "Random Data": """Generate 2000 random numbers between 1 and 10000 in a table with 8 columns per row.""", - - "Text Content": """Generate a comprehensive guide about machine learning with 50 sections, each containing detailed explanations and examples.""" + "Text Content": """Generate a comprehensive guide about how to bring a new product to market in 10 sections, each containing detailed explanations and examples.""" } async def main():