System Core
Sacred Orders
Babel Gardens

Babel Gardens

Last Updated: February 25, 2026

Text-to-signal extraction: embeddings, sentiment/signals, and explainable linguistic outputs.

What it does

  • Extracts structured semantic signals from unstructured text (with confidence + traceability)

  • Produces embeddings and model-driven features for downstream Orders (Pattern Weavers, Neural Engine, Vault)

  • Publishes events and exposes APIs via LIVELLO 2 (service), while keeping LIVELLO 1 pure

  • Epistemic layer: Perception (Linguistic Processing / Semantic Grounding)

  • Mandate: turn multilingual text into structured, auditable artifacts (vectors + signals)

  • Outputs: embeddings, sentiment/emotion signals, topic matches, fused vectors, explainability traces

Charter (mandate + non-goals)

Mandate

  • detect language (or validate caller-provided language)
  • produce semantic embeddings (single/batch, multilingual)
  • expose semantic signals (sentiment/emotion + structured traces)
  • classify text into configurable topics (taxonomy)
  • fuse vectors deterministically when required (synthesis)

Non-goals

  • no business decisions (no ranking, no risk scoring)
  • no domain ontology resolution (belongs to Pattern Weavers)
  • no storage governance (belongs to Vault Keepers / Memory Orders)

Interfaces

  • HTTP (LIVELLO 2): services/api_babel_gardens/
    • Embeddings: /v1/embeddings/*
    • Sentiment: /v1/sentiment/*
    • Emotion: /v1/emotion/detect
    • Profiles + cognitive routing: /v1/profiles/*, /v1/events/*, /v1/routing/*
  • Cognitive Bus (optional, LIVELLO 2): streams listener + event channels

Event contract (Cognitive Bus)

Defined in vitruvyan_core/core/cognitive/babel_gardens/events/__init__.py:

  • babel.embedding.request / babel.embedding.response
  • babel.sentiment.request / babel.sentiment.response
  • babel.emotion.request / babel.emotion.response
  • babel.topic.request / babel.topic.response
  • babel.synthesis.request / babel.synthesis.response
  • babel.error

Code map

  • LIVELLO 1 (pure, no I/O): vitruvyan_core/core/cognitive/babel_gardens/
    • Domain config: domain/config.py (language + cache + taxonomy injection)
    • Topic + language consumers: consumers/classifiers.py
    • Vector fusion consumer: consumers/synthesis.py
    • Events: events/__init__.py
    • Examples: examples/*.yaml (vertical signal schema examples)
  • LIVELLO 2 (service + modules + adapters): services/api_babel_gardens/
    • Embedding engine: modules/embedding_engine.py
    • Sentiment fusion: modules/sentiment_fusion.py
    • Emotion detection: modules/emotion_detector.py
    • Profile processor + cognitive bridge: modules/profile_processor.py, modules/cognitive_bridge.py
    • HTTP routes: api/routes_*.py

Pipeline (happy path)

  1. Service receives text (+ optional language hint)
  2. Language is detected/validated (multilingual-safe path)
  3. Embedding is generated (and optionally cached)
  4. Optional: sentiment/emotion signals are extracted
  5. Optional: vectors are fused (synthesis) for downstream consumers
  6. Result is returned via HTTP and/or emitted on the bus

Consumers (LIVELLO 1)

LanguageDetectorConsumer — language detection (heuristic)

  • File: vitruvyan_core/core/cognitive/babel_gardens/consumers/classifiers.py
  • Purpose: lightweight language detection to support domain purity (adapters can use stronger ML cascades)

TopicClassifierConsumer — taxonomy classification

  • File: vitruvyan_core/core/cognitive/babel_gardens/consumers/classifiers.py
  • Purpose: classify text into topics using a deploy-time taxonomy (YAML-driven, domain-agnostic)

SynthesisConsumer — vector fusion

  • File: vitruvyan_core/core/cognitive/babel_gardens/consumers/synthesis.py
  • Purpose: fuse semantic + sentiment vectors deterministically (semantic_garden_fusion default)

Verticalization (finance pilot)

Finance binds Babel Gardens by providing:

  • topic taxonomy (keywords/categories) via YAML injection
  • signal schema examples (domain-agnostic primitives): vitruvyan_core/core/cognitive/babel_gardens/examples/signals_finance.yaml
  • optional service plugins: services/api_babel_gardens/plugins/finance_signals.py

Rule: the schema/config lives in the vertical, not in the Babel Gardens core.

Comprehension Engine v3 (February 25, 2026)

With BABEL_COMPREHENSION_V3=1, Babel Gardens gains a unified comprehension pipeline that produces both semantic signals and ontology structure in a single LLM call.

Architecture

LayerComponentResponsibility
L1ComprehensionConsumerParse LLM JSON → ComprehensionResult (ontology + semantics)
L2Domain models (FinBERT, etc.)Produce SignalEvidence[] via ISignalContributor
L3SignalFusionConsumerWeighted/bayesian/LLM-arbitrated fusion → FusionResult

New endpoints

  • POST /v2/comprehend — full comprehension (ontology + semantics in one call)
  • POST /v2/fuse — multi-source signal fusion

Consumers (LIVELLO 1)

ComprehensionConsumer — LLM JSON parsing

  • File: vitruvyan_core/core/cognitive/babel_gardens/consumers/comprehension_consumer.py
  • Assembles prompt from domain plugin (ontology + semantics sections)
  • Returns ComprehensionResult with separate OntologyPayload and SemanticPayload

SignalFusionConsumer — signal fusion

  • File: vitruvyan_core/core/cognitive/babel_gardens/consumers/signal_fusion_consumer.py
  • Strategies: weighted (default), bayesian, llm_arbitrated
  • Fuses L1 (LLM) + L2 (domain model) signals into FusionResult

Registries (LIVELLO 1)

ComprehensionPluginRegistry + SignalContributorRegistry

  • File: vitruvyan_core/core/cognitive/babel_gardens/governance/signal_registry.py
  • Domain plugins (IComprehensionPlugin) shape both ontology and semantics prompts
  • Signal contributors (ISignalContributor) add domain-calibrated signals (lazy-loaded)

Adapters (LIVELLO 2)

  • ComprehensionAdapter (services/api_babel_gardens/adapters/comprehension_adapter.py) — LLM orchestration
  • SignalFusionAdapter (services/api_babel_gardens/adapters/signal_fusion_adapter.py) — fusion + LLM arbitration

Graph integration

  • comprehension_node (core/orchestration/langgraph/node/comprehension_node.py) replaces both pattern_weavers_node and emotion_detector_node with full backward compatibility

Plugin system

  • IComprehensionPlugin — shapes both ontology and semantics prompt sections
  • ISignalContributor — contributes domain-specific signals to fusion
  • Built-in: GenericComprehensionPlugin (domain-agnostic)
  • Finance: FinanceComprehensionPlugin (11 entity types, FinBERT signals) + FinBERTContributor

Contracts

  • contracts/comprehension.py — ComprehensionResult, SemanticPayload, IComprehensionPlugin, ISignalContributor, SignalEvidence, FusionResult
  • contracts/pattern_weavers.py — OntologyPayload (reused, PW-owned)

Tests

  • 49 core tests (tests/test_comprehension_engine.py): contracts, consumers, registries, cross-domain
  • 29 finance tests (tests/test_finance_comprehension_plugin.py): plugin, FinBERT, fusion
  • All 78 pass ✅

For the full architectural rationale, see Semantic & Ontology Architecture.

Memory OrdersPattern Weavers