Skip to main content

IntentMapper

What it is​

  • A small intent-matching utility that:
    • Indexes a list of Intent objects in an in-memory VectorStore using embeddings.
    • Retrieves the most similar intents to an input string (intent text or raw user prompt).
  • It also initializes an LLM client (ChatOpenAI), but the current map_prompt implementation does not call the LLM.

Public API​

Enums​

  • IntentScope
    • DIRECT, ALL
  • IntentType
    • AGENT, TOOL, RAW

Dataclass​

  • Intent
    • Fields:
      • intent_value: str - canonical text for the intent (used for embedding/indexing).
      • intent_type: IntentType - classification of the intent target.
      • intent_target: Any - arbitrary target object associated with the intent.
      • intent_scope: Optional[IntentScope] = IntentScope.ALL - scope metadata.

Class: IntentMapper​

  • __init__(intents: list[Intent], embedding_model: Embeddings | None = None)
    • Builds a VectorStore and indexes intent.intent_value for all provided intents.
    • If embedding_model is provided, uses it for document/query embeddings.
    • If embedding_model is None, logs a warning and instantiates OpenAIEmbeddings(model="text-embedding-3-large").
    • Defers VectorStore creation until embeddings are computed, then infers vector dimension from the first vector.
    • Initializes self.model = ChatOpenAI(model="gpt-4.1-mini").
    • Sets a system_prompt string (not used by current mapping methods).
  • get_intent_from_value(value: str) -> Intent | None
    • Returns the Intent whose intent_value exactly matches value, else None.
  • map_intent(intent: str, k: int = 1) -> list[dict]
    • Performs vector similarity search over indexed intents for the given text.
    • Augments each result dict with result["intent"] containing the matched Intent object.
  • map_prompt(prompt: str, k: int = 1) -> Tuple[list[dict], list[dict]]
    • Maps the raw prompt directly via map_intent.
    • Returns a tuple: ([], prompt_results) (first element always empty in current implementation).

Configuration/Dependencies​

  • Embeddings/vector store dependencies:
    • Optional custom embedding_model implementing LangChain Embeddings
    • OpenAIEmbeddings for default fallback behavior
    • VectorStore with:
      • add_texts(texts, embeddings=..., metadatas=...)
      • similarity_search(query_embedding, k=...)

Usage​

from naas_abi_core.services.agent.beta.IntentMapper import Intent, IntentMapper, IntentType

intents = [
    Intent(intent_value="write a report", intent_type=IntentType.AGENT, intent_target="report_agent"),
    Intent(intent_value="calculate an arithmetic result", intent_type=IntentType.TOOL, intent_target="calculator"),
]

mapper = IntentMapper(intents)

results = mapper.map_intent("I need to write a report about AI trends", k=1)
top = results[0]["intent"]
print(top.intent_value)     # e.g. "write a report"
print(top.intent_type)      # IntentType.AGENT
print(top.intent_target)    # "report_agent"

_, prompt_results = mapper.map_prompt("3 / 4 + 5", k=1)
print(prompt_results[0]["intent"].intent_value)

Caveats​

  • If embedding_model is not provided, a warning is logged and OpenAIEmbeddings(model="text-embedding-3-large") is used.
  • The internal VectorStore is initialized only after embeddings are generated, and its dimension is inferred from the first vector.
  • map_prompt does not use the configured LLM; it only does embedding similarity against intent_value strings and returns ([], prompt_results).
  • Returned items from map_intent/map_prompt are dicts whose base shape is determined by VectorStore.similarity_search; this module only adds an "intent" key.