PowerOn/wiki
3
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

rev ai process

Browse source
This commit is contained in:
ValueOn AG 2025-12-01 19:16:01 +01:00
parent c898175855
commit ab8a182001
1 changed files with 82 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

82
appdoc/doc_concept_callAiContent_flow.md Normal file
View file

@ -0,0 +1,82 @@
# callAiContent Process Flow
## Function: `callAiContent(prompt, options, contentParts, outputFormat, title, parentOperationId)`
### High-Level Flow
```
START callAiContent
│
├─> Initialize operation tracking (aiOperationId)
├─> Determine operationType (if not set)
│
├─> [BRANCH 1] IMAGE_GENERATE → Handle image generation → RETURN
│
├─> [BRANCH 2] WEB_SEARCH/WEB_CRAWL → Handle web operations → RETURN
│
└─> [BRANCH 3] Unified Document Generation Path (outputFormat always set, defaults to "txt")
│
├─> IF contentParts provided:
│ │
│ ├─> SEPARATE parts:
│ │ ├─> imageParts = [typeGroup=="image" OR mimeType starts "image/"]
│ │ ├─> textParts = [typeGroup in ["text","table","structure"] OR mimeType starts "text/"]
│ │ └─> otherParts = [rest] → SKIP
│ │
│ ├─> [PART 2.1] PROCESS IMAGES:
│ │ FOR each imagePart:
│ │ ├─> Create AiCallRequest with USER PROMPT + IMAGE_ANALYSE
│ │ ├─> Call aiObjects.call() → vision model extracts data
│ │ └─> Replace imagePart with text ContentPart (extracted data)
│ │
│ ├─> [PART 2.2] ADD TEXT PARTS as-is
│ │
│ └─> Convert processedParts to text string → content_for_generation
│
├─> Build generation prompt (includes content_for_generation)
├─> Call _callAiWithLooping for JSON generation
├─> Parse JSON and render to outputFormat (txt, docx, xlsx, pdf, etc.)
└─> RETURN AiResponse with rendered document
```
### Detailed Flow: Unified Document Generation Path (lines 1055-1257)
```
IF outputFormat specified:
│
├─> IF contentParts provided:
│ │
│ ├─> SEPARATE parts (same logic as text path):
│ │ ├─> imageParts, textParts, otherParts
│ │
│ ├─> [PART 2.1] PROCESS IMAGES:
│ │ └─> Same as text path: vision models with USER PROMPT
│ │
│ ├─> [PART 2.2] ADD TEXT PARTS as-is
│ │
│ └─> Convert processedParts to text:
│ └─> content_for_generation = "\n\n".join([...])
│
├─> Build generation prompt:
│ └─> generation_prompt = buildGenerationPrompt(outputFormat, prompt, title, content_for_generation, None)
│
├─> Call _callAiWithLooping(generation_prompt, ...) for JSON generation
├─> Parse generated JSON
├─> Render to outputFormat (xlsx, docx, pdf, etc.)
└─> RETURN AiResponse with rendered document
```
### Key Points
1. **Unified Path**: All formats (txt, docx, xlsx, pdf, etc.) use the same document generation path. If `outputFormat` is not specified, it defaults to "txt".
2. **User Prompt is Critical**: The prompt from `ai.process` (user's intention) is used to extract data from images, not a generic "extract all text" prompt.
3. **Vision Models**: Images are processed with `IMAGE_ANALYSE` operation type, which routes to vision-capable models.
4. **Text Extraction**: After vision processing, images are replaced with text ContentParts containing the extracted data.
5. **JSON Structure**: All formats generate JSON structure first, then render to the target format (txt rendering extracts text from JSON structure).
6. **Binary Skipping**: Binary and other non-text parts are excluded from processing.