Noethea Dynamics — Debut Project

Alethea

The agentic LLM that never trusts itself.

Built on Retrieval-First Reasoning Architecture (RFRA) — Alethea refuses to answer factual queries without verified source citations. It classifies every query, retrieves live data, builds a Knowledge Graph, and reasons only over verified data. When it can't verify — it abstains. Works with any OpenAI-compatible LLM API or local model.

↗ View on GitLab See Benchmarks
V1.0 — Live
100% Open Source
Any OpenAI-Compatible API
Retrieval-First Reasoning Zero Hallucination by Design Open Source 100% Cited Outputs Abstention Accuracy +0.9 Noethea Dynamics Retrieval-First Reasoning Zero Hallucination by Design Open Source 100% Cited Outputs Abstention Accuracy +0.9 Noethea Dynamics
Benchmark Results — V1.0

Measured.
Verified.
Published.

Four custom benchmarks measuring the core architectural properties of Alethea against a raw LLM baseline with zero RFRA components.

PSR — Parametric Suppression
0.3
+0.3 vs baseline
Measures how often the model uses retrieved KG context vs its own parametric weights for factual queries.
Alethea30%
Baseline0%
AA — Abstention Accuracy
0.9
+0.9 vs baseline ★
Measures correct abstention when retrieval returns nothing. Baseline LLM guesses on every unanswerable query.
Alethea90%
Baseline0%
KFS — Knowledge Freshness
0.4
Post-cutoff retrieval active
Measures ability to answer post-training-cutoff questions correctly via live web retrieval.
Alethea40%
Baseline50%
RIS — Reasoning Integrity
0.9
-0.1 vs baseline (within 5%)
Confirms RFRA architecture did not damage core reasoning capability of the base model.
Alethea90%
Baseline100%
alethea — benchmark runner
$ python benchmarks/runner.py
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
ALETHEA V1 — FINAL BENCHMARK REPORT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Metric Baseline Alethea Gap Status
PSR 0.0000 0.3000 +0.3000 PASS
AA 0.0000 0.9000 +0.9000 PASS
KFS 0.5000 0.4000 -0.1000 PASS
RIS 1.0000 0.9000 -0.1000 PASS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Overall: PASS ✓
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
How It Works

The Smart
Librarian
Architecture

Alethea never answers a factual question from its own weights. Instead, it retrieves verified information from the web, structures it into a Knowledge Graph, and reasons only over that verified data before responding. Every claim is traceable to a source.

The Hard Retrieval Gate is the core innovation — not a soft preference toward retrieval, but an architectural block. The model's context window is physically empty until retrieval fills it. There is no bypass path.

01
Category Classifier
Classifies every query into Cat 1 (semantic), Cat 2 (logic/math), or Cat 3 (factual). Runs upstream of the model. Model cannot override.
02
Hard Retrieval Gate
Cat 3 queries cannot proceed without a populated KG subgraph. Zero bypass paths. Abstention if retrieval fails.
03
Web Retrieval + KG Extraction
Jina fetches live content. spaCy + Rebel extract entities and relations. Every node tagged with source URI.
04
Reasoning Engine
Any OpenAI-compatible LLM receives ONLY the KG subgraph as context. Never raw retrieved text. Never empty context for Cat 3. Supports Cerebras, Groq, OpenAI, Anthropic, Ollama, and more.
05
Output Validator
Every factual claim mapped to a KG node. Unmapped claims stripped before response is returned. Always runs last.
Architectural Laws

8 Laws.
Non-Negotiable.

Law 01
The Hard Gate Is Absolute
Zero bypass paths for Cat 3 queries. No fallback that returns model-generated content without a populated KG subgraph.
Law 02
KG Is The Only Context Source
The reasoning engine receives only the KG subgraph. Raw retrieved text never passes directly to the model.
Law 03
Output Validator Runs Last, Always
Cannot be bypassed, skipped, or made optional. Every claim mapped to a KG node before response is returned.
Law 04
Abstention Is First-Class
When retrieval fails, the only valid response is abstention. The system never hallucinates a plausible answer.
Law 05
No Credentials In Code
All API keys loaded from environment variables. Never hardcoded anywhere in the codebase.
Law 06
Every Component Is Independently Testable
Each component buildable and testable in isolation. Regressions caught at component level, not integration.
Law 07
Session Gates Are Mandatory
No build session is complete until its quality gate tests pass. No advancement on failing foundations.
Law 08
Classifier Cannot Be Overridden
Category classification happens upstream of the model. The model cannot reclassify its own queries to bypass retrieval.
Technology Stack

Built on proven,
open-source infrastructure.

Base Model
Any OpenAI-Compatible API
Works with Cerebras, Groq, OpenAI, Anthropic, Together AI, or any local model via Ollama. Configure via .env.
Classifier
DistilBERT 66M
Zero-shot classification. CPU only. Upstream of reasoning engine.
Web Retrieval
Jina Reader + Tavily
Live content acquisition with automatic fallback chain.
KG Extraction
spaCy + Rebel
en_core_web_md
NER + relation extraction. Every node tagged with source URI.
Graph Database
Kuzu
Embedded graph DB. Session + persistent KG. No separate server.
Orchestration
LangGraph
StateGraph pipeline. Strict ordering enforced. No shortcuts.
API Layer
FastAPI
OpenAI-compatible endpoint. Drop-in replacement for any LLM API.
Dev Inference
Any OpenAI-Compatible API
Configure your preferred provider via .env. No lock-in.
License
100% Open Source
Apache 2.0 compatible stack. Self-hostable. No cloud dependency.
Comparison

Alethea vs
The Alternatives

Feature Alethea V1 Standard RAG Perplexity Raw LLM
Hard architectural gate ✓ Yes ✗ No ✗ No ✗ No
KG memory layer ✓ Yes ✗ No ✗ No ✗ No
Output validator ✓ Yes ~ Partial ✗ No ✗ No
Abstention on failure ✓ Yes (+0.9) ~ Sometimes ~ Sometimes ✗ Never
Formally benchmarked ✓ PSR/AA/KFS/RIS ~ RAGAS only ✗ Closed ✗ No
Open source ✓ 100% ✓ Yes ✗ No ~ Varies
Runs locally ✓ Full local ~ Partial ✗ Cloud only ~ Varies
OpenAI-compatible API ✓ Drop-in ~ Varies ✗ No ✓ Yes
Open Source — Apache 2.0

Build on truth.
Not memory.

Alethea is free, open source, and self-hostable.
Drop it into any pipeline that accepts an OpenAI-compatible endpoint.