Configuration Reference

OneXMonitor parameters

Parameter	Type	Description
`endpoint`	string	Ingestion base URL. Determines which platform environment receives your logs (sandbox, development, production). Fetch environment-specific endpoints from the OneX Observability Dashboard.
`api_key`	string	API key for authentication (from the same dashboard).
`enable_logging`	bool	When `true`, enables SDK logging at INFO level (framework detection, signal export, etc.). Use for local debugging. Default: `false`.
`sample_rate`	float	Fraction of requests to export (0.0–1.0). `1.0` = 100%, `0.1` = 10%. Reduces volume in high-traffic production. Default: `1.0`.
`max_requests_per_minute`	int \| None	Cap on batch export requests per minute to avoid overloading the platform. `None` = no cap. Default: `None`.

Config options

Key	Type	Default	Description
`hidden_state_sample_tokens`	int	4	Number of tokens sampled per hidden-state payload.
`hidden_state_sample_features`	int	32	Number of features sampled per token.
`hidden_state_precision`	int	6	Decimal precision for rounded tensor samples (use -1 to disable rounding).
`capture_full_hidden_state`	bool	false	Set `true` to include the entire hidden-state tensor (use cautiously).
`payload_sample_items`	int	5	Max items serialized per list/dict when exporting request payloads.
`payload_tensor_sample`	int	32	Flat element sample size for tensor summaries in request payloads.
`payload_max_depth`	int	2	Maximum depth for nested request payload serialization.
`request_metadata`	dict	`{}`	Extra metadata appended to every request payload/response event.
`request_payload_endpoint`	string	derived	Override for the payload ingestion endpoint (defaults to `/api/requests/payload`).
`request_response_endpoint`	string	derived	Override for the response ingestion endpoint (defaults to `/api/requests/response`).
`capture_request_payload`	bool	true	When `false`, request payloads (model inputs and raw context) are not sent to the platform. Neural signals still go to `/api/signals/batch`.
`capture_response_payload`	bool	true	When `false`, response payloads (model outputs and application responses) are not sent. Neural signals still go to `/api/signals/batch`.
`enable_logging`	bool	false	When `true`, enables SDK logging at INFO level for debugging. Can also be passed as a top-level parameter to `OneXMonitor`.
`sample_rate`	float	1.0	Fraction of requests to export (0.0–1.0). Per-request sampling; can also be passed as a top-level parameter.
`max_requests_per_minute`	int \| None	None	Cap on batch export requests per minute. Can also be passed as a top-level parameter.
`capture_logits`	bool	false	When `true`, model logits are extracted from the forward output and included in the response payload (shape + sample). For Uncertainty, Calibration, and Confidence assessments.
`capture_probabilities`	bool	false	When `true`, probabilities (softmax of logits) are computed and included in the response payload (shape + sample). Requires logits to be present in the model output.
`logits_sample_size`	int	64	Max number of values to include in the logits/probabilities sample (flattened). Limits payload size.
`signal_queue_maxsize`	int	10000	Max neural signals queued in memory before the exporter drops new ones (`queue.Full`). LLMs emit many signals per token; increase if you see warnings while the batch API is healthy. Capped at 1,000,000.
`request_queue_maxsize`	int	2000	Max queued request payload/response export events. Capped at 100,000.
`capture_llm_extension`	bool	true	PyTorch only. When `true`, each exported signal includes an optional `llm_extension` object with extra LLM context (generation phase, output summaries, model metadata). Set `false` to reduce payload size.
`llm_extension_cuda_memory`	bool	true	PyTorch only. When `true` and CUDA is available, `llm_extension` may include GPU memory usage summaries.
`capture_llm_representation`	bool	true	PyTorch. One `representation.<policy>` signal per forward with an embedding vector for OOD-style assessments.
`llm_representation_policy`	string	last_token	`last_token`, `prompt_last` (second-to-last token when seq>1), or `mean_pool`.
`llm_representation_max_dims`	int	512	Max dimensions exported in the representation embedding list.
`capture_generation_context`	bool	true	Adds `phase` (prefill/decode), `sequence_length`, and related fields to `llm_extension` / representation.
`capture_logits_extras`	bool	true	Adds margin, perplexity proxy, and top-1 probability summaries on output signals.
`capture_final_norm`	bool	true	Causal LMs. Export summaries from the final normalization layer before the language-model head.
`capture_kv_cache_meta`	bool	true	Include KV cache shape metadata in LLM output summaries (no raw tensor dumps).

LLM extension and export queues

For decoder / LLM workloads, the PyTorch adapter can emit multiple signals per forward pass. The background exporter sends them in batches. If the HTTP path is slower than signal production, in-process queues can fill.

signal_queue_maxsize / request_queue_maxsize: Increase these when running large models or long generations and you observe exporter warnings about a full queue.
capture_llm_extension: Extra structured metadata for backends and debugging. Disable if you do not need it.
llm_extension_cuda_memory: Disable if you want to avoid reading CUDA memory statistics in llm_extension.

Example:

monitor = OneXMonitor(
    api_key="your-api-key",
    endpoint="onex-ingestion-endpoint",
    config={
        "signal_queue_maxsize": 50_000,
        "request_queue_maxsize": 5_000,
        "capture_llm_extension": True,
        "llm_extension_cuda_memory": True,
    },
)

Sampling and throughput control

Sample rate: When sample_rate is less than 1.0, the SDK exports only a fraction of requests. The decision is made once per request (by request_id), so either all signals for a request are exported or none. Use this in high-traffic production to reduce volume while keeping representative coverage (e.g. sample_rate=0.1 for 10% of requests).
Max requests per minute: When set, the exporter limits how many batch HTTP requests it sends to the platform per 60-second sliding window. Batches are sent as soon as possible under this cap. Use this to avoid overloading the ingestion API when many models or processes export concurrently.

Example:

monitor = OneXMonitor(
    api_key="your-api-key",
    endpoint="onex-ingestion-endpoint",
    sample_rate=0.1,                  # export 10% of requests
    max_requests_per_minute=120,      # cap at 120 batch requests per minute
)

Logits and probabilities capture

For Uncertainty, Calibration, and Confidence assessments, the platform can use model logits and/or probabilities. By default they are not captured (privacy and payload size).

capture_logits: When true, the SDK extracts the logits tensor from the model forward output (e.g. HuggingFace CausalLMOutputWithPast.logits) and adds a logits field to the response payload with shape, sample (first N values), and numel.
capture_probabilities: When true, the SDK computes softmax(logits) and adds a probabilities field with the same structure. Requires the model output to expose logits (e.g. .logits attribute or first element of a tuple).
logits_sample_size: Caps how many values are included in the sample (default 64). Use a higher value for small classification heads; keep lower for large vocabularies.

When either option is enabled, the main output field in the response is serialized without the raw logits tensor so you avoid duplicating data and control size via the sampled fields. Example:

monitor = OneXMonitor(
    api_key="your-api-key",
    endpoint="onex-ingestion-endpoint",
    config={
        "capture_logits": True,
        "capture_probabilities": True,
        "logits_sample_size": 128,
    },
)

Start Here

Models

AI Agents

Reference

OneXMonitor parameters

Config options

LLM extension and export queues

Sampling and throughput control

Logits and probabilities capture

​OneXMonitor parameters

​Config options

​LLM extension and export queues

​Sampling and throughput control

​Logits and probabilities capture

OneXMonitor parameters

Config options

LLM extension and export queues

Sampling and throughput control

Logits and probabilities capture