LiteLLM

Connect LiteLLM (library or proxy) to Radicas — native gen_ai telemetry via its built-in OTel callback.

Status: experimental — implemented (best-effort activation per the LiteLLM docs), not yet fixture-verified; no registry card yet. LiteLLM is a router/proxy, not an agent framework: you get LLM-request spans only, no agent/tool tree — still enough for cost estimation, and an ideal capture point for opaque clients routed through the proxy.

Install

pip install "radicas[litellm]"     # = radicas + litellm

Use (library mode)

import radicas

radicas.init(service="router", feature="shared-llm")
# instrument="auto" detects litellm, registers the "otel" callback and sets
# LITELLM_OTEL_V2=true (one gen_ai-compliant trace per request).

import litellm
response = litellm.completion(model="claude-sonnet-4-5", messages=[...])
radicas.shutdown()

Proxy mode: the callback is configured server-side in the proxy's own config (litellm_settings: callbacks: ["otel"]) with the standard OTEL_EXPORTER_OTLP_* env — the SDK strategy applies to library usage in your process.

What it emits

Scope name: litellm — expected, pending fixture.
Namespace: native gen_ai.* (with LITELLM_OTEL_V2=true), including its own gen_ai.cost.* estimates.
Coverage: expected, pending fixture — model + token usage documented as emitted; cache tokens depend on the upstream provider.

Shim behavior

Client-side adapter: none expected (gen_ai.* native) — a light card may still be added for the radicas.source_framework stamp and any name drift.
Server-side OTTL: pending the same card.

Known limitations

LLM-request spans only: no invoke_agent root, no execute_tool — the tree views stay empty for this source; the LLM/cost views populate.
LiteLLM's gen_ai.cost.* is an estimate feeding reconciliation, not a replacement for authoritative invoice data.
Content capture uses LiteLLM's own kill switches (turn_off_message_logging) in addition to the SDK's opt-in.

Verify it works

SELECT model, input_tokens, output_tokens, estimated_cost_usd
  FROM otel.llm_call ORDER BY ts DESC LIMIT 5;

Bench recipe: providers/litellm/python in radicas-integrations (recipe only — the runnable example is part of the pending deep-dive).