extractors

Consolidated lightweight feature-extraction service for the DiDi AI platform (LOT 1, Modul 2 — Extractoare ML). Hosts the CPU-only / LLM-delegating feature extractors that do not warrant a dedicated GPU model server.

Heavy model services stay separate: deepfake/BusterX → video-analysis, speech-to-text/Whisper → audio. They require GPU at runtime and are not folded in here.

Capabilities

Endpoint	Status	Type	Notes
`POST /v1/metadata`	✅ implemented	CPU, deterministic	EXIF, video metadata (ffprobe), spectrogram, ELA, integrity/signature
`POST /v1/sentiment`	✅ implemented	LLM-delegated	LLM gateway + Romanian-aware prompt (irony/sarcasm)
`POST /v1/ocr`	✅ implemented	LLM vision	verbatim text from images via the vision model
`POST /v1/ner`	✅ implemented	model (`.[ml]`)	GLiNER multilingual, configurable entity types, GPU-capable
`POST /v1/detect`	✅ implemented	model (`.[ml]`)	YOLO object detection (COCO), boxes+conf, GPU-capable

The metadata extractors are selected automatically from the detected media type (image / video / audio).

Prerequisites

Python 3.10+
ffmpeg/ffprobe on PATH (provided by the container image) — required for video metadata and non-WAV audio decoding. Image extractors (EXIF/ELA/ integrity) work without it.

Run (local)

uv sync
uv run uvicorn extractors.app:app --port 54400

Test

uv run pytest                      # local (needs deps)
./deploy/deploy.sh test            # or build the wheel in Docker

In the container image, ffmpeg-dependent paths run for real; the unit tests are fixture-driven so they pass without the binaries.

Deploy

cp .env.example .env               # set EXTRACTORS_TAG / PORT / MAX_UPLOAD_MB
./deploy/deploy.sh up

Ports follow the datacenter convention: 54400 (dev), 14400 (prod).

See API.md for the full request/response contract.

2 KiB Raw Blame History