[Enhancement] Add LangChain-compatible DocumentLoader for audio transcription

[deepgram-robot]

Summary

Add a DeepgramAudioLoader class that implements the LangChain BaseLoader interface, enabling developers to load audio files as LangChain Document objects with Deepgram transcription — ready for use in RAG pipelines, agent tools, and chain workflows.

Problem it solves

Developers using LangChain for AI agent and RAG workflows currently have no built-in way to ingest audio content via Deepgram. They must write custom glue code to transcribe audio and convert results into LangChain Document format. A native loader in the SDK (or as a companion utility) would make Deepgram a first-class audio source in the LangChain ecosystem, matching patterns developers already use for PDF, CSV, and web loaders.

Proposed API

from deepgram.integrations.langchain import DeepgramAudioLoader

loader = DeepgramAudioLoader(
    file_paths=["meeting.mp3", "call.wav"],
    api_key="...",  # or from env
    model="nova-3",
    smart_format=True,
    diarize=True,
)

documents = loader.load()
# Each Document has:
#   page_content = transcript text
#   metadata = {"source": "meeting.mp3", "duration": 120.5, "speakers": 3, ...}

Acceptance criteria

• Implements LangChain BaseLoader interface (or BaseBlobParser)
• Supports file paths, URLs, and raw bytes as input
• Populates Document metadata with Deepgram response fields (duration, speakers, confidence, model)
• Works with both sync and async LangChain patterns
• Documented with usage example
• Compatible with existing SDK API

---

Raised by the DX intelligence system.