The Llama Index integration enables to monitor your RAG pipelines with a single line of code:
LlamaIndex instrumentation
literalai_client.instrument_llamaindex()
The Llama Index integration already support LLM tracing. You should not use it in conjunction with other LLM provider integrations such as OpenAI.
The Llama Index integration in the Python SDK is compatible with Llama Index starting with version 0.10.58.
LlamaIndex offers a variety of concepts to interact with LLMs:
  • Query Engines
  • LLMs
  • Agents
We explain what each concept leads to in terms of Thread, Run and Generation logs, and show you a visual of what you can expect on Literal AI.

​
Query Engines

from literalai import LiteralClient
from llama_index.core import Document, VectorStoreIndex

client = LiteralClient()
# The magic happens here
client.instrument_llamaindex()

index = VectorStoreIndex.from_documents([Document.example()])
query_engine = index.as_query_engine()

# Here we are outside of a thread so the integration will create one for us
query_engine.query("Tell me about LLMs")

# If we call LlamaIndex from inside a context, the integration will use it
with client.thread(name="Test LlamaIndex Thread") as thread:
    query_engine.query("Tell me about LLMs")

# If we make several calls inside the context, they will all be logged in the same thread
with client.thread(name="LlamaIndex Questions") as thread:
    query_engine.query("Tell me about LLMs")
    query_engine.query("What is RAG ?")

# Optional: make sure to flush the events before stopping the client
client.flush_and_stop()
Each Thread will result in the following tree on Literal AI :
A Llamaindex RAG thread on Literal AI

A LlamaIndex RAG thread on Literal AI

​
LLMs

LlamaIndex offers wrappers around LLM providers to interact with their APIs.

​
llm.chat

The methods llm.chat and llm.stream_chat both generate a standalone Generation:
A LlamaIndex LLM call - Standalone Generation

A LlamaIndex LLM call - Standalone Generation

Please note that LlamaIndex token usage is not available for streaming methods due to limitations in the event data present on the LLMChatEndEvent for chunk completions.However, the Literal AI platform defaults token counts computation to the cl100k_base tokenizer which is a fair approximation of the expected token usage.

​
llm.predict_and_call

The llm.predict_and_call also results in a standalone Generation on the Literal AI platform. Specifically, LlamaIndex does not trigger events related to tool calls and we recommend decorating your tools’ function definitions with @literalai_client.step(type="tool", name="My Tool") to view the calls performed. Note that a Step of type tool cannot be standalone on the Literal AI platform and we thus recommend you to add a contextual Step wrapper around your llm.predict_and_call call, as such:
LlamaIndex llm.predict_and_call
with literalai_client.step(type="run", name="Predict & Call")
    llm.predict_and_call(...)

​
Agents

LlamaIndex has the concept of agent as part of its FunctionCallingAgent and specifically derived an OpenAIAgent with specificities to the OpenAI model offerings. Function calling agents can be tuned in a variety of ways, but the general idea is that they iteratively perform the configured LLM calls with tool options until the LLM deems it unnecessary to call a tool. When calling an agent.chat, you can expect to obtain a “run” Step of the following form:
A LlamaIndex agent chat - Agent Run with multiple intermediate steps

A LlamaIndex agent chat - Agent Run with multiple intermediate steps

The tool calls in the stack above show only because the functions themselves are decorated with an @literalai_client.step(type="tool", name="My Tool").