- System
- Pipelines & Workflows
- Codebase Maps
- System Overview
- Celery Queues
The full stack in one diagram: SvelteKit SPA, Nginx, FastAPI backend, Celery workers, Postgres, OpenSearch, MinIO, and the LLM provider. Start here.
The queues behind "Celery Workers": gpu, cpu, nlp, embedding, redaction, utility, download, and cloud-asr/cpu-transcribe, which the pipeline plan picks per file, plus what lite runs without the gpu workers.
- Transcription Pipeline
- Post-Transcription Enrichment
- Cloud ASR, Speakers & Voiceprints
- Speaker Clustering & Naming
- Search & Indexing
- RAG / Chat Retrieval
Upload, the per-file pipeline plan, transcription on the queue it picks, speaker detection inline or in a post stage, voiceprints on this server, indexing and delivery.
What happens after a transcript is finalized: voiceprinting, gender/age detection, LLM speaker-ID suggestions, summarization, topic extraction, analytics, and the per-file record of what ran where.
How cloud transcription gets its speakers (the vendor itself, pyannote.ai, or another provider through the adapter), how timing decides the merge, and why voiceprints are still computed on this server.
From a raw voiceprint to automatic profile matching, cross-file clustering, human-reviewed promotion to a named speaker, and the unassign rules that keep a correction from being undone.
The query path (BM25 + kNN fused via RRF, then quarantine/redaction filtering) and the write path (chunking + embedding) in one diagram.
Query → scope/plan → retrieval → the local-vs-remote redaction guard → LLM → output redaction → streamed answer with citations.
- Backend Module Map
- Frontend Module Map
How backend/app/{api,auth,schemas,services,tasks,models,db,core,utils} actually relate: one services/ layer called by both the API and Celery.
How the SvelteKit frontend's routes/, components/, stores/, and lib/ fit together, including the community/cloud seam.