Advanced algorithmsAvailable algorithmsllm

llm

The llm module provides a function to send text to a large language model (LLM) and return the completion directly from Cypher. It uses LiteLLM for model-agnostic completion, so you can use OpenAI, Anthropic, Ollama, and other providers with the same interface.

Source code
TraitValue
Module typeutil
ImplementationPython
Parallelismsequential

Requirement: Install LiteLLM in the MAGE Python environment: pip install litellm.
Set the API key for your provider (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY). For local providers like Ollama, you can set api_base in the config or LITELLM_API_BASE in the environment.

Functions

complete()

Sends text to an LLM and returns the completion as a string (e.g. for summarization or text generation).

Input:

  • text: str ➡ Input text to send to the LLM (e.g. concatenated node text to summarize or complete).
  • config: Map (OPTIONAL) ➡ Configuration map. Defaults to {}.

Config options:

NameTypeDescription
modelstringModel name (e.g. ollama/llama2, openai/gpt-4o-mini). Falls back to env LITELLM_MODEL or a default.
api_basestringBase URL for the API (e.g. http://localhost:11434 for Ollama). Can also be set via LITELLM_API_BASE.
system_promptstringSystem prompt for the completion. Default: "Complete the following text."

Output:

  • Returns str ➡ The model’s completion (trimmed string). Empty string if the model returns no content.

Usage:

Using default or environment configuration:

RETURN llm.complete("Summarize in one sentence: Memgraph is a graph database.");

With explicit config (e.g. Ollama):

RETURN llm.complete("Hello", {model: "ollama/llama2", api_base: "http://localhost:11434"});

With a custom system prompt:

RETURN llm.complete(
  "List the main topics.",
  {system_prompt: "You are a concise assistant. Reply with a short bullet list."}
);

Using graph data (e.g. summarize node text):

MATCH (n:Article)
WITH collect(n.title + ": " + n.abstract) AS texts
WITH reduce(s = "", t IN texts | s + t + "\n") AS combined
RETURN llm.complete(combined, {system_prompt: "Summarize the following articles in 2 sentences."}) AS summary;
leiden_community_detectionllm_util