AI data flow

What enters external LLM providers from Sophia and what gets persisted.

flowchart LR
  subgraph upload ["Resume"]
    S3[(S3)]
    Extract["Text extract"]
    LLM1["Parse LLM"]
    Cand[(Candidate)]
  end
  subgraph live ["Interview"]
    Bot["External bot"]
    Int[(Transcript)]
  end
  subgraph report ["Report"]
    SQS["Report queue"]
    LLM2["Report LLM"]
    Rep[(InterviewReport)]
  end
  S3 --> Extract --> LLM1 --> Cand
  Cand --> Bot --> Int
  Int --> LLM2
  Cand --> LLM2
  SQS --> LLM2 --> Rep

• Resume parse — PDF/DOCX → plain text locally; LLM sees text only (not raw bytes). Response is ephemeral until the user saves a candidate; resume.text is stored from deterministic extraction at save.
• Schedule — Copies resume.text into the interview snapshot; no LLM at schedule time.
• Live interview — Bot writes transcript and proctoring; dialogue LLM is outside Sophia.
• Report — Enqueued after first successful credit settlement; model input is snapshot context, resume excerpt (up to 4k chars), and indexed transcript.
• Deterministic reports — Very short interviews or insufficient candidate answers skip the LLM and store a fixed insufficient-evidence result.
• Bot cost metrics — Optional PATCH from the bot; Sophia stores telemetry, does not recompute LLM spend.