Experimental Language and Cognition

Scope

Grilly includes an experimental cognitive stack centered around vector symbolic representations and fast ingestion.

Primary components

  • experimental.language: WordEncoder, SentenceEncoder, InstantLanguage, SVC ingestion pipeline.

  • experimental.cognitive: WorldModel, CognitiveController, simulation and understanding helpers.

  • experimental.vsa: binary/holographic ops and GPU-backed VSA acceleration.

Instant language model

InstantLanguage provides:

  • on-the-fly word encoding

  • sentence memory

  • template learning

  • similarity search over stored sentence vectors

  • SVC data ingestion with optional GPU acceleration through SVCIngestionEngine

Cognitive controller

CognitiveController orchestrates:

  1. language understanding

  2. world-model consistency checks

  3. candidate generation and simulation

  4. confidence-gated response selection

Minimal ingestion example

from grilly.experimental.language.svc_loader import load_svc_entries_from_dicts
from grilly.experimental.cognitive.controller import CognitiveController

entries = load_svc_entries_from_dicts([
    {
        "id": "x0",
        "text": "Gravity attracts mass.",
        "svc": {"s": "Gravity", "v": "attracts", "c": "mass"},
        "pos": ["NOUN", "VERB", "NOUN", "PUNCT"],
        "deps": ["nsubj", "ROOT", "dobj", "punct"],
        "lemmas": ["gravity", "attract", "mass", "."],
        "root_verb": "attract",
        "realm": "science",
        "source": "manual",
        "complexity": 0.3,
    }
])

controller = CognitiveController(dim=1024)
result = controller.ingest_svc(entries, verbose=False)
print(result.summary())

Operational guidance

  • Use smaller dimensions during iteration (dim=512 or 1024) for faster tests.

  • Turn off expensive n-gram paths for large ingestion runs when needed.

  • Persist ingestion state with checkpoint utilities for resumable workflows.