CatyAI Architecture Glossary
Version: v3.0 | Date: June 2026
Terms are listed alphabetically. Cross-references in bold.
AKL V2 (AI Knowledge Layer V2)
The per-widget response layer that wraps every outbound answer in a signed, structured envelope consumable by AI agents. AKL V2 is isolated from V1 — it uses Ed25519 (RFC 8032) via @noble/ed25519 and RFC 8785 JCS canonicalization, whereas V1 used Node's native crypto module.
Outputs: answer, sources, content_hash, ttl, available_actions, signature, jwks_uri.
Code: src/services/akl/response-builder.js, src/services/sentinel-v2/signing.js
Cameleon Lens
The vertical detection and payload normalization layer. Converts raw, platform-specific payloads (Shopify, WooCommerce, GoMag, WhatsApp messages) into a UnifiedSemanticContext struct. Named "Cameleon" because it adapts to any input shape without requiring callers to know the source platform.
Outputs: { vertical, entity_type, primary_id, primary_name, attributes, metadata }
Code: src/core/cameleon/schema.js, src/core/cameleon/adapters/
content_hash
A SHA-256 digest of the tuple { query, answer, source_count } included in every AKL V2 response. Allows agents to verify response integrity independently of the EdDSA signature.
ECC (Elliptic Curve Cryptography)
The cryptographic family used by CatyAI for offer signing and transaction validation. The specific curve is Ed25519 (Curve25519 / Edwards curve), accessed via the @noble/ed25519 library. ECC is chosen over RSA for its shorter key length (32 bytes seed vs. 2048-bit RSA) and faster signing at equivalent security levels.
EdDSA / Ed25519
The signing algorithm used in NAP V3 and AKL V2. Ed25519 produces 64-byte deterministic signatures from a 32-byte private seed. Keys are stored in AWS Secrets Manager as { kid, seed_hex, public_hex }. Verification uses the JWKS endpoint (/.well-known/jwks.json).
Standard: RFC 8032 (EdDSA), RFC 8037 (JWK representation)
GEO Gateway
The public, auth-free API layer designed for AI crawlers. Exposes POST /geo/v2/answer (signed knowledge queries), GET /geo/v2/llms.txt (NAP V3 document), and catalog/manifest endpoints. "GEO" stands for Generative Engine Optimization — the AI-era equivalent of SEO.
Code: src/api/geo/
Intent TTL (Time to Live)
The expiry window attached to each market intent injected into Qdrant metadata. Values range from 12–48 hours, determined by the ttl_hours field returned by DeepSeek. After expiry, the metadata reverts to the base product vectors, keeping the latent space clean.
JCS (JSON Canonicalization Scheme)
RFC 8785 — a deterministic JSON serialization used before signing. Ensures that two JSON objects with the same data but different key ordering produce the same byte sequence, making signatures stable across serializers.
Package: canonicalize npm package
Market Intelligence Worker
A cron-driven ECS service that runs at 04:00 UTC and every 4 hours, coordinating the Predictive Intent Engine cycle. Also handles daily competitor monitoring, algorithm change detection, and NAP anchor enrichment. Entry point for runPredictiveIntentCycle().
Code: src/workers/market-intelligence-worker.js
NAP V3 (Native AI Protocol V3)
CatyAI's machine-readable discovery protocol for AI crawlers. Combines three public endpoints — llms.txt, /.well-known/ai-context, and GET /geo/v2/manifest — to expose verified business data without requiring HTML rendering. V3 adds EdDSA signing to every offer, enabling Zero-Trust transactions.
Note: Internal signing code uses the label "NAP V2" (referring to the signing algorithm version). External-facing documentation uses "NAP V3" (the protocol version).
O(K) Complexity
The scalability property of the Predictive Intent Engine. K = the fixed number of market verticals (e.g., 10–20 industries). Regardless of how many merchant widgets are active (N = 4,500+), the Scout runs only K LLM queries per cycle, not N. Per-query cost is effectively O(1) since the intent vectors are pre-computed.
Payload Hot-Update
The Qdrant operation that injects dynamic intent metadata into existing product vector records without re-embedding the product itself. Uses Qdrant's metadata upsert API. Paired with Intent TTL to auto-expire stale enrichments.
Code: autoseo-intent-worker.js → hotUpdateQdrantMetadata()
Predictive Intent Engine (The Scout)
The asynchronous market intelligence subsystem. A Scout worker queries DeepSeek every 4 hours for top-3 buying intent signals per industry, fans them out via AWS SNS, processes them through SQS into the AutoSEO Intent Worker, and injects them as TTL-scoped metadata into Qdrant. Reduces query-time complexity from O(N) to O(K).
Code: src/services/predictive-intent.js
Semiotic Architecture
CatyAI's foundational data model. Contrasted with semantic architecture (stores product definitions, answers "what is this?"), semiotic architecture intercepts real-time market signals and anomalies (answers "what does the market intend right now?"). The term comes from semiotics — the study of signs and signification — applied to the idea that market behavior carries predictive signals beyond static meaning.
Example: A viral social trend associating "rose quartz" with "exam anxiety relief" is a semiotic signal. Semantic architecture misses it; CatyAI's semiotic layer surfaces it and adjusts vector weights within hours.
SNS Fan-out
The AWS messaging pattern used to distribute market intent events to all subscribers simultaneously. CatyAI uses a single FIFO topic (caty-intent-global.fifo) with industry encoded in the message payload. New industries require zero infrastructure changes.
TTL Intent
A market intent enrichment with an explicit expiry. Stored as Qdrant metadata fields intent_text, urgency_score, intent_expires_at. When intent_expires_at is in the past, the vector search falls back to base product semantics.
UnifiedSemanticContext
The normalized output struct produced by the Cameleon Lens. All downstream services (AKL V2, brain, SAG agents) consume this shape, never the raw platform payload. Decouples platform-specific parsing from business logic.
Fields: vertical, entity_type, primary_id, primary_name, attributes, metadata.source_platform, metadata.normalized_at, metadata.detection_method
Vector Hot-Update
See Payload Hot-Update.
Zero-Trust Transaction
A purchase executed by an AI agent that is mathematically validated without relying on a human-visible HTML frontend. The agent receives an EdDSA-signed product offer, verifies the signature via JWKS, and submits it back for fulfillment. The backend rejects any transaction with an invalid or missing signature, making price hallucination impossible at the infrastructure level.
Named after the Zero-Trust network security model (never trust, always verify), applied to AI-to-AI commerce.
Cross-reference: See Architecture KB for implementation details and code references.