gateway/modules/services/serviceAi/subStructureGeneration.py

335 lines
15 KiB
Python

# Copyright (c) 2025 Patrick Motsch
# All rights reserved.
"""
Structure Generation Module
Handles document structure generation, including:
- Generating document structure with sections
- Building structure prompts
"""
import json
import logging
from typing import Dict, Any, List, Optional
from modules.datamodels.datamodelExtraction import ContentPart
from modules.datamodels.datamodelAi import AiCallOptions, OperationTypeEnum, PriorityEnum, ProcessingModeEnum
logger = logging.getLogger(__name__)
class StructureGenerator:
"""Handles document structure generation."""
def __init__(self, services, aiService):
"""Initialize StructureGenerator with service center and AI service access."""
self.services = services
self.aiService = aiService
def _getUserLanguage(self) -> str:
"""Get user language for document generation"""
try:
if self.services:
# Prefer detected language if available (from user intention analysis)
if hasattr(self.services, 'currentUserLanguage') and self.services.currentUserLanguage:
return self.services.currentUserLanguage
# Fallback to user's preferred language
elif hasattr(self.services, 'user') and self.services.user and hasattr(self.services.user, 'language'):
return self.services.user.language
except Exception:
pass
return 'en' # Default fallback
async def generateStructure(
self,
userPrompt: str,
contentParts: List[ContentPart],
outputFormat: str,
parentOperationId: str
) -> Dict[str, Any]:
"""
Phase 5C: Generiert Chapter-Struktur (Table of Contents).
Definiert für jedes Chapter:
- Level, Title
- contentParts (unified object with instruction and/or caption per part)
- generationHint
Args:
userPrompt: User-Anfrage
contentParts: Alle vorbereiteten ContentParts mit Metadaten
outputFormat: Ziel-Format (html, docx, pdf, etc.)
parentOperationId: Parent Operation-ID für ChatLog-Hierarchie
Returns:
Struktur-Dict mit documents und chapters (nicht sections!)
"""
# Erstelle Operation-ID für Struktur-Generierung
structureOperationId = f"{parentOperationId}_structure_generation"
# Starte ChatLog mit Parent-Referenz
self.services.chat.progressLogStart(
structureOperationId,
"Chapter Structure Generation",
"Structure",
f"Generating chapter structure for {outputFormat}",
parentOperationId=parentOperationId
)
try:
# Baue Chapter-Struktur-Prompt mit Content-Index
structurePrompt = self._buildChapterStructurePrompt(
userPrompt=userPrompt,
contentParts=contentParts,
outputFormat=outputFormat
)
# AI-Call für Chapter-Struktur-Generierung mit Looping-Unterstützung
# Use _callAiWithLooping instead of callAiPlanning to support continuation if response is cut
options = AiCallOptions(
operationType=OperationTypeEnum.DATA_GENERATE,
priority=PriorityEnum.QUALITY,
processingMode=ProcessingModeEnum.DETAILED,
compressPrompt=False,
compressContext=False,
resultFormat="json"
)
# Create prompt builder for continuation support
async def buildChapterStructurePromptWithContinuation(
continuationContext: Optional[Dict[str, Any]] = None,
**kwargs
) -> str:
"""Build chapter structure prompt with optional continuation context."""
basePrompt = self._buildChapterStructurePrompt(
userPrompt=userPrompt,
contentParts=contentParts,
outputFormat=outputFormat
)
if continuationContext:
# Add continuation instructions
deliveredSummary = continuationContext.get("delivered_summary", "")
elementBeforeCutoff = continuationContext.get("element_before_cutoff", "")
cutOffElement = continuationContext.get("cut_off_element", "")
continuationText = f"{deliveredSummary}\n\n"
continuationText += "⚠️ CONTINUATION: Response was cut off. Generate ONLY the remaining content that comes AFTER the reference elements below.\n\n"
if elementBeforeCutoff:
continuationText += "# REFERENCE: Last complete element (already delivered - DO NOT repeat):\n"
continuationText += f"{elementBeforeCutoff}\n\n"
if cutOffElement:
continuationText += "# REFERENCE: Incomplete element (cut off here - DO NOT repeat):\n"
continuationText += f"{cutOffElement}\n\n"
continuationText += "⚠️ CRITICAL: The elements above are REFERENCE ONLY. They are already delivered.\n"
continuationText += "Generate ONLY what comes AFTER these elements. DO NOT regenerate the entire JSON structure.\n"
continuationText += "Start directly with the next chapter that should follow.\n\n"
return f"""{basePrompt}
{continuationText}
Continue generating the remaining chapters now.
"""
else:
return basePrompt
# Call AI with looping support
# NOTE: Do NOT pass contentParts here - we only need metadata for structure generation
# The contentParts metadata is already included in the prompt (contentPartsIndex)
# Actual content extraction happens later during section generation
aiResponseJson = await self.aiService.callAiWithLooping(
prompt=structurePrompt,
options=options,
debugPrefix="chapter_structure_generation",
promptBuilder=buildChapterStructurePromptWithContinuation,
promptArgs={
"userPrompt": userPrompt,
"outputFormat": outputFormat,
"services": self.services
},
useCaseId="chapter_structure", # REQUIRED: Explicit use case ID
operationId=structureOperationId,
userPrompt=userPrompt,
contentParts=None # Do not pass ContentParts - only metadata needed, not content extraction
)
# Parse the complete JSON response (looping system already handles completion)
extractedJson = self.services.utils.jsonExtractString(aiResponseJson)
parsedJson, parseError, cleanedJson = self.services.utils.jsonTryParse(extractedJson)
if parseError is not None:
# Even with looping, try repair as fallback
logger.warning(f"JSON parsing failed after looping: {str(parseError)}. Attempting repair...")
from modules.shared import jsonUtils
repairedJson = jsonUtils.repairBrokenJson(extractedJson)
if repairedJson:
parsedJson, parseError, _ = self.services.utils.jsonTryParse(json.dumps(repairedJson))
if parseError is None:
logger.info("Successfully repaired and parsed JSON structure after looping")
structure = parsedJson
else:
logger.error(f"Failed to parse repaired JSON: {str(parseError)}")
raise ValueError(f"Failed to parse JSON structure after repair: {str(parseError)}")
else:
logger.error(f"Failed to repair JSON. Parse error: {str(parseError)}")
logger.error(f"Cleaned JSON preview (first 500 chars): {cleanedJson[:500]}")
raise ValueError(f"Failed to parse JSON structure: {str(parseError)}")
else:
structure = parsedJson
# ChatLog abschließen
self.services.chat.progressLogFinish(structureOperationId, True)
return structure
except Exception as e:
self.services.chat.progressLogFinish(structureOperationId, False)
logger.error(f"Error in generateStructure: {str(e)}")
raise
def _buildChapterStructurePrompt(
self,
userPrompt: str,
contentParts: List[ContentPart],
outputFormat: str
) -> str:
"""Baue Prompt für Chapter-Struktur-Generierung."""
# Baue ContentParts-Index - filtere leere Parts heraus
contentPartsIndex = ""
validParts = []
filteredParts = []
for part in contentParts:
contentFormat = part.metadata.get("contentFormat", "unknown")
# WICHTIG: Reference Parts haben absichtlich leere Daten - immer einschließen
if contentFormat == "reference":
validParts.append(part)
logger.debug(f"Including reference ContentPart {part.id} (intentionally empty data)")
continue
# Überspringe leere Parts (keine Daten oder nur Container ohne Inhalt)
# ABER: Reference Parts wurden bereits oben behandelt
if not part.data or (isinstance(part.data, str) and len(part.data.strip()) == 0):
# Überspringe Container-Parts ohne Daten
if part.typeGroup == "container" and not part.data:
filteredParts.append((part.id, "container without data"))
continue
# Überspringe andere leere Parts (aber nicht Reference, die wurden bereits behandelt)
if not part.data:
filteredParts.append((part.id, f"no data (format: {contentFormat})"))
continue
validParts.append(part)
logger.debug(f"Including ContentPart {part.id}: format={contentFormat}, type={part.typeGroup}, dataLength={len(str(part.data)) if part.data else 0}")
if filteredParts:
logger.debug(f"Filtered out {len(filteredParts)} empty ContentParts: {filteredParts}")
logger.info(f"Building structure prompt with {len(validParts)} valid ContentParts (from {len(contentParts)} total)")
# Baue Index nur für gültige Parts
for i, part in enumerate(validParts, 1):
contentFormat = part.metadata.get("contentFormat", "unknown")
originalFileName = part.metadata.get('originalFileName', 'N/A')
contentPartsIndex += f"\n{i}. ContentPart ID: {part.id}\n"
contentPartsIndex += f" Format: {contentFormat}\n"
contentPartsIndex += f" Type: {part.typeGroup}\n"
contentPartsIndex += f" MIME Type: {part.mimeType or 'N/A'}\n"
contentPartsIndex += f" Source: {part.metadata.get('documentId', 'unknown')}\n"
contentPartsIndex += f" Original file name: {originalFileName}\n"
contentPartsIndex += f" Usage hint: {part.metadata.get('usageHint', 'N/A')}\n"
if not contentPartsIndex:
contentPartsIndex = "\n(No content parts available)"
# Get language from services (user intention analysis)
language = self._getUserLanguage()
logger.debug(f"Using language from services (user intention analysis) for structure generation: {language}")
prompt = f"""CRITICAL OUTPUT REQUIREMENT: This is a PLANNING task, not a generation task. You MUST return EXACTLY ONE complete JSON object. Do NOT generate multiple JSON objects, alternatives, or variations. Do NOT use separators like "---" between JSON objects. Return the single best JSON structure that matches the requirements below.
USER REQUEST (for context):
```
{userPrompt}
```
LANGUAGE: Generate all content in {language.upper()} language. All text, titles, headings, paragraphs, and content must be written in {language.upper()}.
AVAILABLE CONTENT PARTS:
{contentPartsIndex}
TASK: Generate Chapter Structure for the documents to be generated.
IMPORTANT - CHAPTER INDEPENDENCE:
- Each chapter is independent and self-contained
- One chapter does NOT have information about another chapter
- Each chapter must provide its own context and be understandable alone
CONTENT ASSIGNMENT:
- Assign ContentParts to chapters via contentParts object
- For data extraction, the type of a contentPart (image, text, etc.) is NOT relevant - only what is specified in the instruction matters
- Include ALL relevant parts from same source when needed for structured data extraction
- Each contentPart can have either:
- "instruction": For AI extraction prompts (how to process/extract from this part)
- "caption": For user-facing presentation (how to display/reference this part in the document)
- Both can be present if needed
- Chapters without contentParts can only generate generic content (not document-specific)
FORMATTING:
- Formatting is handled automatically - focus on CONTENT and STRUCTURE only
CHAPTER STRUCTURE:
- chapter id, level (1, 2, 3, etc.), title
- contentParts: {{"partId": {{"instruction": "..."}} or {{"caption": "..."}} or both}} - Compact mapping of part IDs to their extraction instructions and/or presentation captions
- generationHint: Self-contained description (if contentParts is empty, must be VERY DETAILED)
RETURN JSON:
{{
"metadata": {{
"title": "Document Title",
"language": "{language}"
}},
"documents": [{{
"id": "doc_1",
"title": "Document Title",
"filename": "document.{outputFormat}",
"chapters": [
{{
"id": "chapter_1",
"level": 1,
"title": "Introduction",
"contentParts": {{
"part_ext_1": {{
"instruction": "Use full extracted text"
}},
"part_img_1": {{
"instruction": "Analyze image for additional details"
}},
"part_img_2": {{
"instruction": "Analyze image for additional details",
"caption": "Figure 1: Overview diagram"
}}
}},
"generationHint": "Create introduction section",
"sections": []
}},
{{
"id": "chapter_2",
"level": 1,
"title": "Main Title",
"contentParts": {{}},
"generationHint": "Create [specific content description] with [formatting details]. Include [required information]. Purpose: [explanation of what this chapter provides].",
"sections": []
}}
]
}}]
}}
OUTPUT FORMAT: Start with {{ and end with }}. Do NOT use markdown code fences (```json). Do NOT add explanatory text before or after the JSON. Return ONLY the JSON object itself.
"""
return prompt