> ## Documentation Index > Fetch the complete documentation index at: https://docs.omnia-voice.com/llms.txt > Use this file to discover all available pages before exploring further. # Request logging > Opt-in capture of your chat traffic — browse it, export it, turn it into fine-tuning datasets. Request logging keeps a **full-content record of your chat completions**, prompts and responses, browsable in the dashboard and exportable as JSONL. It's the raw material for improving your models: review real traffic, build evaluation sets, and curate fine-tuning datasets from what your users actually ask. It is **off by default**. Omnia's standing posture is content-free: we never retain prompts or completions unless you explicitly choose it. Request logging is one of exactly two opt-in exceptions (the other is [response caching](/billing/caching)). ## Enabling it In the [dashboard](https://platform.omnia-voice.com/dashboard) → **Settings → Workspace → Request logging** (owner/admin only): Takes effect at the gateway within a few seconds. 7, 30, or 90 days. Every logged request carries its own deletion deadline; retention is enforced row by row, automatically. ## What gets stored | | Logged? | | --------------------------------------------------------- | ------------------------------------------------ | | Successful chat completions (streaming and non-streaming) | ✅ full messages + response, including tool calls | | Cache-hit replays | ✅ flagged as cached | | Fallback-served responses | ✅ with the model it fell back from | | Errors, partial or dropped streams | ❌ never | | Embeddings / rerank / responses endpoints | ❌ not in v1 | | Requests to dedicated endpoints | ❌ | | Oversized exchanges (> 512 KB) | ❌ skipped, never truncated | Each entry also records the model, [tag and trace id](/reference/observability#request-labels-tags-and-trace-ids), token counts, and finish reason: the signals you filter by when curating. **Credential-shaped strings are scrubbed before storage**: API keys, bearer tokens, and similar patterns in your prompts are replaced with `[redacted]` automatically. This is defense-in-depth, not a guarantee; avoid sending secrets to models in the first place. ## Per-request opt-out With logging enabled, any individual call can stay out: ```bash theme={null} curl https://gateway.omnia-voice.com/v1/chat/completions \ -H "Authorization: Bearer $OMNIA_API_KEY" \ -H "X-Omnia-Log: off" \ ... ``` The header can only **reduce** logging: with the workspace toggle off it does nothing. ## Browsing and exporting The **Logs** page in the dashboard lists your logged traffic newest-first, filterable by model, [tag](#tagging-your-traffic), auto-detected [**segment**](/concepts/segments) (the system-prompt family a request belongs to: zero setup, works retroactively), time range, finish reason, and cache hits. Click any row for the full conversation, message by message, tool calls included. **Export JSONL** downloads the current filter as chat-format lines, one `{"messages": [...]}` object per exchange, with the assistant's reply as the final message. That's the standard shape for fine-tuning and evaluation pipelines, ours or anyone else's. Exports are capped at 10,000 exchanges per download. **Create dataset** curates the current filter into a managed dataset instead: errored, truncated, empty, and duplicate exchanges are dropped automatically and every drop is reported by reason. The dataset keeps its provenance: including the [segment](/concepts/segments) it was cut from, so you always know which application surface trained on it. Optionally hold out 10% as a second, **disjoint** `-eval` dataset. No example appears in both, which is what makes it a valid set to [evaluate](/reference/evals) against later. A dataset is a **frozen snapshot**: it does not age out when the logs it was curated from expire. The same asymmetry applies to [grades](/judges/grading) (labels in the API); grades are permanent, while the logged conversations they graded are deleted on your retention window. Curate and grade what matters before it ages out. ### Multi-turn conversations are folded, not repeated A chat application resends the whole history on every turn, so a 3-turn conversation is logged as **three** exchanges whose messages are nested: ``` {"messages": [Q1, A1]} {"messages": [Q1, A1, Q2, A2]} {"messages": [Q1, A1, Q2, A2, Q3, A3]} ``` Training on those as three separate examples would teach the model turn 1 **three times** and the final turn once, an implicit bias toward the start of every conversation. So we reconstruct them into **one** training example with a `weight` on each assistant turn: ```json theme={null} {"messages": [ {"role": "user", "content": "Q1"}, {"role": "assistant", "content": "A1", "weight": 1}, {"role": "user", "content": "Q2"}, {"role": "assistant", "content": "A2", "weight": 1}, {"role": "user", "content": "Q3"}, {"role": "assistant", "content": "A3", "weight": 1} ]} ``` Every turn survives and every turn is learned **exactly once**. The curation summary tells you when this happened ("842 earlier turns folded into 210 conversations"), so a dataset with fewer lines than exchanges is never a mystery. Conversations are reconstructed from the message content itself, so this works without any tagging on your side. Independent steps of an **agent run** are not folded: they aren't continuations of one another, so they stay as separate training examples. ## Tagging your traffic Label each request with the task it belongs to, and every downstream surface (logs, datasets, evals) can be scoped to that one task: ```bash theme={null} curl https://gateway.omnia-voice.com/v1/chat/completions \ -H "Authorization: Bearer $OMNIA_API_KEY" \ -H "X-Omnia-Tag: support-bot" \ ... ``` Tags are content-free labels (≤64 chars). `X-Omnia-Trace-Id` groups the steps of one multi-step agent run. Neither is ever forwarded upstream. Didn't tag? Auto-detected [segments](/concepts/segments) group logged traffic by system-prompt family with zero setup, retroactively. The tag stays the deliberate override. The two compose, so you can filter by both at once. ## From the API Everything on the Logs page is available to a workspace API key: ```bash theme={null} # Browse (filters: model, tag, segment, finish_reason, cache_hit, start, end, limit, offset) curl "https://gateway.omnia-voice.com/v1/logs?tag=support-bot&limit=25" \ -H "Authorization: Bearer $OMNIA_API_KEY" # Export the same filter as JSONL, straight into your own pipeline curl "https://gateway.omnia-voice.com/v1/logs/export?tag=support-bot" \ -H "Authorization: Bearer $OMNIA_API_KEY" > support-bot.jsonl # Curate it into a training set + a disjoint 10% eval holdout curl -X POST https://gateway.omnia-voice.com/v1/datasets/from_logs \ -H "Authorization: Bearer $OMNIA_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "name": "support-bot-v1", "holdout_pct": 10, "filters": { "tag": "support-bot", "cache_hit": false } }' ``` Every returned row carries its `segment` (the auto-detected [system-prompt family](/concepts/segments)) and `segment_exact` (the exact prompt variant within it); pass a row's `segment` back as `?segment=`, or as `filters.segment` on `POST /v1/datasets/from_logs`, to slice traffic you never tagged. These endpoints return **409** while request logging is off, because an empty list would read as "no traffic" when the truth is "nothing is being captured". See the [Logs & datasets API reference](/api-reference/introduction). ## Privacy posture * Logged content is **workspace-scoped**: stored under your workspace's key, never visible to any other tenant, never used by Omnia for anything. * Retention is **per-row**: each entry is deleted when its window expires, regardless of later settings changes. * Turning the toggle off stops all writes immediately; existing entries simply age out. * Telemetry and observability remain **content-free** exactly as before; this feature adds a separate, consented store; it doesn't change the default.