Struct ProviderTelemetry 

pub struct ProviderTelemetry {

    pub source: String,
    pub server_total_ms: Option<u64>,
    pub server_load_ms: Option<u64>,
    pub server_prompt_eval_ms: Option<u64>,
    pub server_generation_ms: Option<u64>,
    pub server_prompt_tokens: Option<i64>,
    pub server_output_tokens: Option<i64>,
    pub client_wall_ms: Option<u64>,
    pub runtime_context_length: Option<u64>,
    pub runtime_loaded_model: Option<String>,
    pub runtime_memory_bytes: Option<u64>,
    pub runtime_memory_vram_bytes: Option<u64>,
    pub runtime_keep_alive_until: Option<String>,
    pub request_id: Option<String>,
}

Provider-side timing and runtime accounting captured per LLM call.

All fields default to None / empty. Producers fill in what they can extract and leave the rest absent; consumers must treat missing fields as “not reported by this provider”, not “zero”.

Fields

source: String

Wire format the values came from (ollama_chat, openai_usage, …). See source for the canonical strings. Empty when no telemetry was captured.

server_total_ms: Option<u64>

Total server-side wall clock (Ollama total_duration).

server_load_ms: Option<u64>

Time the server spent loading/warming the model (Ollama load_duration). Useful for detecting cold-start latency.

server_prompt_eval_ms: Option<u64>

Prompt-prefill time (Ollama prompt_eval_duration). Anything else would be marketing — this is the field evals key on for prefill regression detection.

server_generation_ms: Option<u64>

Generation/decode time (Ollama eval_duration).

server_prompt_tokens: Option<i64>

Tokens the server reports it prefilled (Ollama prompt_eval_count). Distinct from LlmResult::input_tokens because hosted providers frequently bill different token boundaries than the on-device tokenizer reports.

server_output_tokens: Option<i64>

Tokens the server reports it generated (Ollama eval_count).

client_wall_ms: Option<u64>

Client-side wall clock around the HTTP request. Includes network and streaming latency the server-side counters omit. Recorded for every call regardless of provider.

runtime_context_length: Option<u64>

Context window the model was loaded with (where the runtime reports it; /api/ps for Ollama).

runtime_loaded_model: Option<String>

Exact model id the server resolved. May differ from LlmResult::model when an alias / digest is rewritten upstream.

runtime_memory_bytes: Option<u64>

Total resident bytes for the loaded model (Ollama /api/ps.size).

runtime_memory_vram_bytes: Option<u64>

VRAM bytes for the loaded model (Ollama /api/ps.size_vram).

runtime_keep_alive_until: Option<String>

When the server will unload the model (Ollama /api/ps.expires_at).

request_id: Option<String>

Provider-supplied request id (x-request-id / request_id).

Implementations

impl ProviderTelemetry

pub fn new(source: &str) -> Self

pub fn is_empty(&self) -> bool

Returns true when no meaningful telemetry was captured. A bare client_wall_ms is still “meaningful” — it lets evals reason about per-call latency even for providers that report nothing else.

pub fn ns_to_ms(ns: u64) -> u64

Convert nanoseconds (Ollama’s reporting unit) to milliseconds with integer rounding. Centralized so every Ollama timing field uses the same conversion and zero-vs-None semantics line up.

pub fn from_ollama_done(frame: &Value, source: &str) -> Self

Extract Ollama-shape telemetry from a done=true chat or generate frame. Returns a populated ProviderTelemetry whose source is the caller-provided wire identifier.

pub fn from_openai_usage(usage: &Value, request_id: Option<&str>) -> Self

Extract OpenAI-shape usage telemetry. Most OpenAI-compatible local runtimes only report counts; llama.cpp’s usage.timings extension is folded in here as well so a single envelope captures both shapes.

pub fn from_anthropic_usage(usage: &Value, request_id: Option<&str>) -> Self

Extract Anthropic-shape usage telemetry. Anthropic only reports input/output (and cache) counts — preserving the request id is the most useful incremental signal beyond what LlmResult already carries.

pub fn merge_ollama_ps(&mut self, ps: &OllamaPsModel)

Merge a /api/ps snapshot of a loaded Ollama model into this telemetry envelope. Only fills in fields that were not already populated so a per-call extraction keeps precedence.

pub fn as_vm_dict(&self) -> Option<VmValue>

Render this envelope into the dictionary shape llm_call returns. Returns None if the envelope is empty so callers can omit the key entirely.

Trait Implementations

impl Clone for ProviderTelemetry

fn clone(&self) -> ProviderTelemetry

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ProviderTelemetry

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ProviderTelemetry

fn default() -> ProviderTelemetry

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for ProviderTelemetry

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl PartialEq for ProviderTelemetry

fn eq(&self, other: &ProviderTelemetry) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for ProviderTelemetry

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for ProviderTelemetry