System One AI + PostgreSQL & Cognitive Semantics

Query unstructured data with
calibrated semantic SQL

Extend relational databases with TypeSafe's Jev model. First-class NOUL, CHOICE, and SCORE primitives with relational pushdown optimization, cognitive syntax, and pluggable semantic engines.

npm install jev-ql

12.2x

Cost reduction via Speculative Fan-out

0.46ms

SHA-256 in-tree cache latency

58.0%

Pushdown filter relational pruning ($0)

0 deps

Pure ES modules & zero external packages

Interactive Query Switchboard

Live in-browser execution. Edit any query below; the real client-side JevQL compiler, AST planner, and execution engine run natively in your browser.

Presets:

ENGINE:

Query Text (Cognitive or SQL) — Editable REPL Engine: Zero-Dep Fast Heuristic (<0.1ms)

Query Element ⟷ Jev API Primitive Mapping

Compiler specification: How human-readable query expressions map 1:1 to TypeSafe System One API contracts.

Compiler & Runtime Mapping Specification

Each query expression translates directly into either a relational pushdown filter or a typed Jev semantic primitive.

Query Element / Syntax	Jev API Primitive	Payload & Semantics	Output Type & Mathematical Range
`? "prompt" > 0.7` `NOUL(col, 'prompt')` `IS_TRUE(col, 'prompt')`	noul (Boolean Gating)	Sends `instructions: "prompt"`. Evaluates evidence and probability of condition.	Calibrated Float: `[0.0, 1.0]` Usable directly in mathematical filters (`> 0.7`) or ordering.
`var = opt1 \| opt2 \| opt3` `CHOICE(col, 'prompt', [opts])`	choice (Categorical)	Sends criteria mapping: `criteria: { opt1: null, opt2: null }`. Evaluates multinomial distribution.	Discrete Label: `"opt1"` (argmax winner) Per-class probabilities accessible via `PROB(choice, 'opt1')`.
`var = lvl1 .. lvl2 .. lvl3` `SCORE(col, 'prompt', [lvls])`	score (Continuous Scalar)	Sends ordered criteria spectrum: `criteria: ["low", "med", "high"]`. Projects text onto 1D continuous semantic vector.	Continuous Float: `[0.0, N-1]` Interpolated score (e.g. `1.85`). Supports `AVG()` and numeric ordering.
`CONFIDENCE(choice_or_score)`	confidence (Metadata)	Computed from distribution entropy and peakedness across options.	Calibrated Float: `[0.0, 1.0]` Enables escalation rules (e.g., hand off to human if `< 0.75`).
`PROB(choice, 'opt1')`	probabilities (Logits)	Extracted from the Jev `probabilities` map for option `'opt1'`.	Float Probability: `[0.0, 1.0]` $$\sum P(\text{opt}_i) = 1.0$$
`source: col = 'open'` `WHERE status = 'open'`	Relational Pushdown	AST pre-filter. Evaluated in deterministic code before invoking any AI model.	Pruned Rows ($0 compute cost, 0 network tokens, 0ms latency).

Architecture Benchmark Matrix

Measured compiler efficiency on PolyAI/banking77 (N=100) vs. published hardware specs and analytical baselines. Reproducible via npm run bench.

Architectural Benchmark: Live Measured vs. Literature & Analytical Baselines

Local compiler, relational pushdown, and cache measured in Node.js on Apple Silicon across 100 golden test cases with ground-truth labels.

Architecture Tier	Evaluation Type	Latency / Row	Generated Tokens	Request Count	Grounding & Methodology
JevQL In-Tree Fast-Path Pure ES2022 CPU engine	Measured Live	0.24 ms / row	0 tokens	0 (in-memory)	100% in-tree execution; zero external network calls
JevQL Cache Hit SHA-256 In-Tree LRU	Measured Live	0.006 ms / row	0 tokens	0 (memory hit)	Sub-millisecond repeat execution (0.27ms total)
Relational Pushdown AST Deterministic Pruning	Measured Live	0.00 ms (pruned)	0 tokens	0 calls ($0)	58 of 100 rows dropped before AI engine is invoked
Speculative Fan-out AST Question Bundling	Measured Live	Bundled 2 questions	0 tokens	42 calls (4.8x saved)	Multi-dimensional bundling cuts requests from 200 to 42
JevQL + JevK5 (Open-Weight) allebee/jevk5 · Apache-2.0	Measured Live	288.7 ms / row (Apple M5 · 13.5ms on H100)	0 tokens	42 passes (0 net)	Single forward-pass SemIf logits; 92.9% Top-1 accuracy (39/42) on Banking77
TypeSafe Jev Cloud TypeSafe System One API	Published Spec	Sub-30ms P95	0 tokens	42 calls	Calibrated System One decision probabilities
Naive SQL + LLM Loop Sequential GPT-4o / Claude 3.5	Analytical Model	~600 ms / call (~60,000 ms total)	~180 tokens / row (18,000 tokens total)	200 separate calls	Autoregressive decoding loop for structured JSON responses
Vector Embeddings Bi-Encoder (text-embedding-3 / MiniLM)	Measured Live	4.15 - 25 ms / call (4.15ms on Apple M5 MPS)	0 tokens	100 calls	Cosine similarity; uncalibrated, no negative condition support

🔬 Methodology & Benchmarking Transparency Notes

Measured Live: Relational pushdown pruning (58/100 rows dropped), speculative fan-out consolidation (42 vs 200 requests), and in-tree LRU caching are measured directly on your machine via npm run bench.
Real JevK5 Neural Inference: Measured live on Apple M5 running open-weight model weights (allebee/jevk5 v0.2 Q8_0 GGUF) via llama.cpp with 8 threads: 288.7 ms / row, 0 generated tokens, and 92.9% Top-1 accuracy (39/42 correct). Datacenter GPUs (NVIDIA H100) run JevK5 at ~13.5 ms/row with CUDA graphs.
Why Decision Models Use 0 Tokens: Models like Jev and JevK5 read classification distributions directly from candidate option logits in a single forward pass without autoregressive generation loops.
Why Generative LLMs are Slower: Autoregressive LLMs generate text token-by-token (100–180 tokens per row for JSON schema), requiring hundreds of sequential forward passes and network roundtrips.

Architecture Blueprint

Minimalist execution topology: In-tree pushdown filtering, single-pass batching, and TypeSafe System One.

Pluggable Code Recipes

Pluggable execution architectures: Open-weight SemIf models (JevK5), TypeSafe cloud, on-device WebGPU, and custom engine plugins.

Pluggable Engines in Node.js & Browser

Configure open-weight JevK5, TypeSafe Jev, on-device WebGPU, LLM structured output, or offline heuristic engines seamlessly:

javascript

import jevql from 'jev-ql';

// 1. Open-Weight Decision Model: JevK5 (allebee/jevk5)
const k5Query = jevql.with({ engine: 'jevk5', model: 'alibiserikbay/JevK5' });
const urgent = await k5Query`
  \${tickets}: status = open
  ? "Immediate outage or security incident?" > 0.6
  team = security | infrastructure | billing
  top 5
`;

// 2. TypeSafe Jev System One Cloud API
const jevQuery = jevql.with({ engine: 'jev', apiKey: process.env.TYPESAFE_API_KEY });

// 3. OpenAI / LLM Structured output adapter
const llmQuery = jevql.with({ engine: 'llm', apiKey: process.env.OPENAI_API_KEY });

Custom Engine Plugins

javascript

import { BaseSemanticEngine, registerEngine, jevql } from 'jev-ql';

class MyLocalEngine extends BaseSemanticEngine {
  async evaluateSingleState(state, questions) {
    // Custom ONNX, vLLM, or local embedding logic
    return {
      q1: { type: 'choice', choice: 'billing', confidence: 0.95 }
    };
  }
}

registerEngine('local_onnx', MyLocalEngine);
const myQuery = jevql.with({ engine: 'local_onnx' });