Make the Q&A system prompt accent-correct PT-BR

Answer Q&A in the selected UI language (EN default)

Align MiniCPM5 v1 QA grounding behavior

Answer dashboard questions with hybrid retrieved context

Index the Q&A corpus as a final pipeline stage

Add Embedder port, corpus chunker, and hybrid retrieval logic

Stream Q&A phases through the dashboard interactor

Tell the Q&A model the asker is the travel agent

Skip extraction for media that already has an artifact

Reuse persisted transcriptions when an export is re-processed

Batch audio transcription and survive transcriber failures

Delete the legacy trip/payment/document model and instructor

Cut the pipeline, dashboard and UI over to MediaArtifacts

Add artifact-based reconciliation, pipeline, identity and summary logic

Run OCR-only images and PDF pages in one ZeroGPU task

Keep customer facts in the Q&A prompt; trim only old messages

Process and surface every media item individually, never in batches

Transcribe audio first and merge transcriptions into messages

Read PDF attachments first; skip webp images

Add OCR-only extraction mode behind VIVAMAIS_OCR_ONLY

Run audio transcription concurrently with vision extraction

Simplify dashboard Q&A: minimal prompt, full history, model-first

Stream structured extraction so tables populate per image

Fix empty-context Q&A crash and long voice-note transcription

Use structured vision extraction and surface uncategorized images

Transcribe all images in one ZeroGPU task to outlive the proxy token

Run all OCR before parsing so GPU forks precede llama.cpp load

Apply ruff format across the repository

Harden OCR parsing with JSON grammar, payer rule, and doc detection

Rewrite pipeline: vision OCR -> MiniCPM5 parsing, remove enrichment

Improve extraction quality with higher resolution, better ordering, and text enrichment

Revert hallucination reduction layers 2-6

Add positive validators to vision output normalization

Reject hallucinated trip locators and placeholder fields

Recover trips with ticket-first vision extraction

Reject schema placeholders and fix upload progress UI

Rewrite vision extraction for real WhatsApp photo patterns

Fix image routing for WhatsApp Android auto-named files

Switch UI to dual local models and restore image routing

Rewrite dashboard search with max context and chunked streaming

Stream CRM snapshots after every image in BuildCrmCard

Add TextModel.stream port and per-token streaming in API adapter

Pivot to CPU-only: transformers_cpu adapter, streaming UI, remove legacy backends

Remove image-routing heuristic — all images sent to vision

Inject Clock into pipeline with time-based rules (B4)

Tighten Repository port types (B3)

Image routing — skip unnecessary vision calls (C2)

Decompose BuildCrmCard god object (B1)

Populate Flight segments from ticket extraction (A5)

Wire payment reconciliation into CRM card with profit (A3)