Skip to main content

EntitytoSPARQLAgent

What it is​

  • A LangGraph/LangChain-based agent that:
    • Extracts entities from an input text using BFO (Basic Formal Ontology) mappings.
    • Fetches possible object properties for the extracted entity classes from a triple store.
    • Generates a SPARQL INSERT DATA statement representing entities and relationships.
    • Saves intermediate artifacts (original text, entities JSON, object properties JSON, SPARQL) to object storage.

Public API​

Module-level constants​

  • NAME: "Entity_to_SPARQL"
  • DESCRIPTION: Agent description string.
  • SYSTEM_PROMPT: Long ontology-engineering prompt used as default system prompt.

Functions​

  • create_agent(agent_shared_state: Optional[AgentSharedState] = None, agent_configuration: Optional[AgentConfiguration] = None) -> Optional[Agent]
    • Factory that constructs and returns an EntitytoSPARQLAgent using the default model (naas_abi.models.default.get_model()), MemorySaver checkpointer, and provided (or new) shared state/configuration.

Classes​

  • EntityExtractionState(ABIAgentState)

    • Agent state extension with:
      • entities: list[dict[str, Any]]
      • object_properties: list[dict[str, Any]]

  • EntitytoSPARQLAgent(Agent)

    • Core agent implementation.
    • Key methods (graph nodes):
      • entity_extract(state: EntityExtractionState) -> Command
        • Calls the chat model with a dedicated extraction prompt and updates messages with the model response (expected to be JSON).
      • prep_data(state: EntityExtractionState) -> Command
        • Parses JSON entities from the last model message.
        • Adds a generated ABI URI (http://ontology.naas.ai/abi/<uuid>) to each entity.
        • Queries object properties per distinct class_uri via GetObjectPropertiesFromClassWorkflow.
        • Saves init_text.txt, entities.json, object_properties.json to object storage.
        • Updates state with entities and object_properties.
      • create_sparql(state: EntityExtractionState) -> Command
        • Calls the chat model to generate a SPARQL INSERT DATA statement from entities, object_properties, and the original message.
        • Saves insert_data.sparql to object storage.
        • Updates messages with an AIMessage containing the SPARQL.
      • call_model(state: EntityExtractionState) -> Command
        • Invokes the tool-capable model (_chat_model_with_tools) on the current messages (optionally prepending state["system_prompt"]).
        • Ends execution (goto="__end__") with the response in messages.
      • build_graph(patcher: Optional[Callable] = None)
        • Builds and compiles the execution graph:
          • START -> entity_extract -> prep_data -> create_sparql -> call_model

Configuration/Dependencies​

  • Requires a configured ABIModule engine and services:
    • SERVICES.triple_store (used by GetObjectPropertiesFromClassWorkflow)
    • SERVICES.object_storage (used by StorageUtils)
  • Uses:
    • langchain_core (chat model + message types)
    • langgraph (StateGraph, Command, checkpointing)
    • naas_abi_core agent framework (Agent, AgentSharedState, AgentConfiguration, ABIAgentState)
    • naas_abi_core.utils.JSON.extract_json_from_completion for parsing model output
  • Storage path:
    • datastore/ontology/entities_to_sparql/<timestamp YYYYMMDDHHMMSS>

Usage​

from langchain_core.messages import AIMessage
from naas_abi.agents.EntitytoSPARQLAgent import create_agent

agent = create_agent()

# Build the internal graph (required before running, depending on Agent base implementation)
agent.build_graph()

# Provide input text as a message in the agent state and run the compiled graph
initial_state = {"messages": [AIMessage(content="Alice signed a contract with Bob yesterday.")]}

# If the base Agent exposes a LangGraph app, it's typically something like:
result = agent.graph.invoke(initial_state)

# The final SPARQL is stored in the last message content
print(result["messages"][-1].content)

Caveats​

  • entity_extract expects the model to return only JSON (per prompt). prep_data will fail if JSON cannot be extracted from the last message.
  • prep_data assumes state["messages"][-2] exists (used to save init_text.txt); the graph order must ensure there is an original message before the extraction response.
  • SPARQL generation is model-driven; correctness depends on model compliance with the prompt constraints.
  • Default constructor arguments use mutable defaults (tools=[], agents=[]), which may cause shared state across instances if those lists are mutated.