PGS Engine — The first retrieval system that tells you what it didn't find

Get Started

Three lines to your first PGS query

            const { PGSEngine } = require('pgs-engine');
const engine = new PGSEngine({ sweepProvider, synthesisProvider });
const result = await engine.execute('What are we missing?', graph);
// result.answer includes cross-domain connections and explicit absences
        

The Problem

Standard RAG doesn't know what it doesn't know

No coverage awareness

Retrieves top-20 from a 586-node graph. The other 566 nodes? Never examined. You don't know what you're missing.

No absence signals

Can't report "I searched for X and it's not here." Silence is indistinguishable from not looking.

No cross-domain discovery

Each query retrieves from one region of embedding space. Connections between distant clusters are invisible.

How It Works

Four phases, full coverage

Phase 0

Partition

Louvain community detection splits graph into topological communities

Phase 1

Route

Cosine similarity ranks partitions by relevance to query

Phase 2

Sweep

Parallel full-fidelity LLM passes per partition

Phase 3

Synthesize

Cross-domain synthesis over all sweep outputs

The sweep contract

Every partition sweep returns a structured report — including what it did not find.

Domain State

What this partition covers relative to the query

Findings

Key discoveries with Node ID citations

Outbound Flags

Specific connections to other partitions

Absences

What was searched for and NOT found

Benchmark

Real data. Real queries. Measured.

3.4%

Standard RAG

20/586 nodes examined

100%

Full PGS

586/586 nodes across 11 partitions

100%+

Layered

Standard context + full PGS deep pass

Standard: 37s | PGS: 81s | Layered: 110s

Full benchmark with response comparison →

The Gold Pattern

Layered search: standard query feeds PGS

Run a standard top-K query first. It establishes what's already known. Then PGS sweeps every partition with that context — finding what standard retrieval missed.

Standard Query

Top-K retrieval, fast context

→

Context Injection

Standard results become prior knowledge

→

PGS Deep Pass

Full sweep finds what's missing

vs. Competitors

What PGS finds that others can't

Feature	Standard RAG	GraphRAG	Mem0	PGS
Coverage awareness	×	Partial	×	✓ with %
Absence reporting	×	×	×	✓
Cross-domain discovery	×	Via summaries	×	Outbound flags
Convergence detection	×	×	×	✓
Session continuity	×	×	✓	✓
Works with any LLM	Via framework	OpenAI only	Via framework	✓
Zero dependencies	×	×	×	✓

Use Cases

Where PGS shines

🔍

Research Synthesis

What do we know about X? What's contradictory? What's missing?

✨

Novelty Detection

What's surprising or unexpected in this knowledge base?

🔎

Gap Analysis

When 4 of 6 partitions report no findings on X, that's a high-confidence gap.

🧭

Deep Exploration

Session tracking tells you exactly what % was examined. Continue until 100%.

🔗

Cross-Domain Discovery

Outbound flags in sweeps find connections that span partition boundaries.

The Deeper Innovation

The knowledge graph architecture

PGS queries graphs built by COSMO — a research system with spreading-activation memory, Hebbian learning, 13 semantic edge types, and state-dependent Watts-Strogatz topology maintenance.

PGS is the query layer. The graph beneath it is what makes it possible.

Explore the Architecture

Provider Agnostic

Bring your own LLM

Zero runtime dependencies. PGS works with any provider through a simple generate() interface.

OpenAI Anthropic xAI Ollama Cloud Any OpenAI-compatible

Start finding what you're missing

PGS is open source, MIT licensed, and ready to use.

GitHub

npm install notforyou23/pgs-engine

The first retrieval system that tells you what it didn't find

586 nodes of conformal field theory research