Documentation Index
Fetch the complete documentation index at: https://docs.trodo.ai/docs/llms.txt
Use this file to discover all available pages before exploring further.
Use this when the provider isn’t auto-instrumented (local Ollama, vLLM, a self-hosted inference server, a partner API). You emit every span yourself.
import trodo from 'trodo-node';
trodo.init({ siteId: process.env.TRODO_SITE_ID });
const LLM_URL = 'http://ollama.internal/api/chat';
export async function answer(userId, question) {
const { result } = await trodo.wrapAgent('raw-http-agent', async (run) => {
run.setInput({ question });
const body = {
model: 'llama3.1:70b',
messages: [{ role: 'user', content: question }],
};
// Option A — withSpan + setLlm. Explicit control over span timing.
const respA = await trodo.withSpan({ kind: 'llm', name: 'ollama.chat' }, async (span) => {
span.setInput(body);
const r = await fetch(LLM_URL, { method: 'POST', body: JSON.stringify(body) }).then((x) => x.json());
span.setLlm({
model: r.model,
provider: 'ollama',
inputTokens: r.prompt_eval_count,
outputTokens: r.eval_count,
});
span.setOutput(r);
return r;
});
// Option B — trackLlmCall. One-shot, less code.
const r = await fetch(LLM_URL, { method: 'POST', body: JSON.stringify(body) }).then((x) => x.json());
await trodo.trackLlmCall({
model: r.model,
provider: 'ollama',
inputTokens: r.prompt_eval_count,
outputTokens: r.eval_count,
prompt: body,
completion: r,
metadata: { endpoint: '/api/chat' },
});
run.setOutput({ answer: r.message?.content });
return r.message?.content;
}, { distinctId: userId });
return result;
}
import trodo, os, httpx
trodo.init(site_id=os.environ["TRODO_SITE_ID"])
LLM_URL = "http://ollama.internal/api/chat"
def answer(user_id, question):
with trodo.wrap_agent("raw-http-agent", distinct_id=user_id) as run:
run.set_input({"question": question})
body = {"model": "llama3.1:70b", "messages": [{"role": "user", "content": question}]}
# Option A — span + set_llm
with trodo.span("ollama.chat", kind="llm") as span:
span.set_input(body)
r = httpx.post(LLM_URL, json=body).json()
span.set_llm(
model=r["model"],
provider="ollama",
input_tokens=r["prompt_eval_count"],
output_tokens=r["eval_count"],
)
span.set_output(r)
# Option B — track_llm_call
r = httpx.post(LLM_URL, json=body).json()
trodo.track_llm_call(
model=r["model"],
provider="ollama",
input_tokens=r["prompt_eval_count"],
output_tokens=r["eval_count"],
prompt=body,
completion=r,
metadata={"endpoint": "/api/chat"},
)
run.set_output({"answer": r["message"]["content"]})
return r["message"]["content"]
Option A vs Option B
Both land identical spans in the database. Pick by style:
| withSpan + setLlm | trackLlmCall |
|---|
| Shape | Wrapper around the call | Call-then-record |
| Errors | Thrown exceptions set status='error' automatically | You have to catch + record yourself |
| Prefer for | The call is the span’s unit of work | You already have the response object |
Custom pricing
If the pricing table doesn’t cover your model — self-hosted, zero-cost, negotiated rate — pass cost explicitly:
await trodo.trackLlmCall({
model: 'llama3.1:70b',
provider: 'ollama-self-hosted',
inputTokens: 1200,
outputTokens: 320,
cost: 0, // self-hosted, don't bill
});
See also
