Sigil Python Framework Module: LiteLLM

sigil-sdk-litellm is a LiteLLM callback handler that exports generation telemetry to Sigil.

Installation

pip install sigil-sdk sigil-sdk-litellm
pip install litellm

Quickstart

import litellm
from sigil_sdk import Client
from sigil_sdk_litellm import SigilLiteLLMLogger

client = Client()
handler = SigilLiteLLMLogger(client=client)

litellm.callbacks = [handler]

response = litellm.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

client.shutdown()

Streaming

import litellm
from sigil_sdk import Client
from sigil_sdk_litellm import SigilLiteLLMLogger

client = Client()
litellm.callbacks = [SigilLiteLLMLogger(client=client)]

response = litellm.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Give me three reliability tips."}],
    stream=True,
)
for chunk in response:
    content = chunk.choices[0].delta.content
    if content:
        print(content, end="", flush=True)
print()

client.shutdown()

Configuration

All options are keyword-only on SigilLiteLLMLogger:

The create_sigil_litellm_logger factory accepts the same parameters.

Per-Request Metadata

The handler resolves agent_name, agent_version, and conversation_id from per-request LiteLLM metadata, falling back to the static values from handler init. This is useful when multiple agents share a single LiteLLM proxy.

response = litellm.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Continue our chat."}],
    metadata={
        "agent_name": "search-agent",
        "agent_version": "v2",
        "conversation_id": "conv-abc-123",
    },
)

For conversation_id, the handler also checks session_id and thread_id metadata keys as fallbacks.

LiteLLM Proxy (Docker)

When running LiteLLM as a proxy server in Docker, register the handler via a callback file next to your config.

1. Extend the Docker image:

FROM ghcr.io/berriai/litellm:v1.82.3-stable.patch.2
RUN pip install sigil-sdk sigil-sdk-litellm

2. Create a callback file (sigil_callback.py, same directory as config.yaml):

import os

from sigil_sdk import Client
from sigil_sdk.config import AuthConfig, ClientConfig, GenerationExportConfig
from sigil_sdk_litellm import SigilLiteLLMLogger

client = Client(ClientConfig(
    generation_export=GenerationExportConfig(
        protocol="http",
        endpoint=os.environ["SIGIL_ENDPOINT"],
        auth=AuthConfig(
            mode="basic",
            tenant_id=os.environ.get("SIGIL_AUTH_TENANT_ID", ""),
            basic_password=os.environ.get("SIGIL_AUTH_TOKEN", ""),
        ),
    ),
))
sigil_handler = SigilLiteLLMLogger(
    client=client,
    agent_name="litellm-proxy",
)

3. Reference it in config.yaml:

model_list:
  - model_name: gpt-4o-mini
    litellm_params:
      model: openai/gpt-4o-mini

litellm_settings:
  callbacks: sigil_callback.sigil_handler

The proxy resolves sigil_callback.sigil_handler by importing sigil_callback.py from the config directory and using the sigil_handler instance.

4. Mount both files and run:

docker run -d \
  -v $(pwd)/config.yaml:/app/config.yaml \
  -v $(pwd)/sigil_callback.py:/app/sigil_callback.py \
  -e SIGIL_ENDPOINT=https://your-sigil-endpoint \
  -e SIGIL_AUTH_TENANT_ID=your-tenant \
  -e SIGIL_AUTH_TOKEN=your-key \
  -p 4000:4000 \
  your-litellm-image \
  --config /app/config.yaml

The callback file reads connection details from environment variables. Adjust the AuthConfig mode to match your deployment (see sigil_sdk.config for tenant, bearer, and basic modes).

Behavior

• Mode mapping: non-stream calls -> SYNC, stream calls -> STREAM with first-token timestamp.
• Provider detection: uses custom_llm_provider from LiteLLM's standard logging object.
• Failed calls are recorded with the error attached via set_call_error.
• Chat completion call types (completion, acompletion, text_completion, atext_completion) are recorded as generations.
• Embedding call types (embedding, aembedding) are recorded as OTel embedding spans (no generation export). The span carries input/token counts and dimensions; the input text is attached only when the handler's capture_inputs is set and the SDK's EmbeddingCaptureConfig.capture_input=True. Embedding spans require a configured OTel tracer.
• Image, audio, and transcription call types are skipped.
• Framework tags are always set:
  • sigil.framework.name=litellm
  • sigil.framework.source=handler
  • sigil.framework.language=python
• LiteLLM request_tags are forwarded as litellm.tag.<value>.
• Token usage includes detailed breakdowns (cached tokens, reasoning tokens) when the provider returns them.
• Tool calls and tool results in messages are mapped to Sigil's tool call/result parts.
• Reasoning/thinking text is captured as THINKING parts, ordered before the assistant text. It is read from thinking_blocks when present (including redacted blocks), otherwise from the flat reasoning_content string.

Call client.shutdown() during teardown to flush buffered telemetry.