Skip to content

Real-world application examples

Release validation should use application-shaped inputs, not only spans assembled from isolated attribute fixtures. The repository includes a vendor-key-free replay of a banking voice-RAG turn:

examples/real_world/voice_rag_replay.py

The replay reads an actual WAV recording and a knowledge-base corpus. It emits one trace containing separate speech-to-text, local embedding, retrieval, and grounded-answer spans. Audio is attached only to the speech span and must be represented as type=audio with media_mime_type=audio/wav.

Run it against the remote collector:

python examples/real_world/voice_rag_replay.py `
  --audio <path-to>/account_balance.wav `
  --corpus <path-to>/kb_corpus.py `
  --transcript "Please check my account balance and recent transactions." `
  --endpoint http://<collector-host>:4318

The example does not claim to call an ASR, embedding, vector database, or LLM vendor. It validates the telemetry path with the application payload shape while remaining runnable without paid credentials. Use deterministic lab fixtures for the WAV and corpus, not production customer data.

Document OCR

The companion replay covers the other multimodal workflow: PDF/image ingestion and structured document extraction.

python examples/real_world/document_ocr_replay.py `
  --document <path-to>/invoice.pdf `
  --endpoint http://<collector-host>:4318

The emitted OCR span must report type=document and media_mime_type=application/pdf for a PDF. An image input must report type=image and its image MIME type. Use --no-export for a local shape-only check. This replay validates file handling and telemetry; it does not claim that a provider-backed OCR model ran.

A full voice-RAG application runs ASR → embedding → vector retrieval → reranking → grounded answer → TTS, and an OCR application runs document bytes through multimodal extraction. These replays cover the telemetry shape of both without the providers.