▶ readme.txt

no-inference

A purely symbolic conversational engine that answers questions, writes essays, solves math problems, roleplays characters, and holds conversations. No neural networks involved. This system cannot hallucinate or lie. It only returns real, factual information, a huge safety improvement over classic LLMs.

Proof of concept Research project

no-inference

⚠

Warning. This is not the best this system could be. We know where the gaps are. The pattern matcher could cover more ground, the knowledge base could be thousands of entries deep, the NLG pipeline could produce more natural prose, and the math solver could handle calculus. But building that takes data, testing, and server time we do not have.

If this project had the resources to run continuous evaluation and iterate on the knowledge base at scale, it would compete with chatbot services that cost millions to run. As it stands, it is a demonstration that the approach works. We think that is worth something.

▶ Demos

Try the demo

This is running on a server with 100mb of RAM.

Chat demo Code editor demo

▶ How it works

How it works

no-inference classifies every query by intent, then routes it to a specialized handler. No neural networks, no GPUs, no model downloads, no API fees.

query → intent detection → handler → knowledge retrieval → response

The best way to have a good conversation is to not use a neural network. -- Ancient proverb, probably

▶ Safety

Safety first

Classic LLMs are trained to predict the next token. When they do not know an answer, they guess, and they guess with full confidence. This is how hallucinations happen. They can lie, make up facts, cite nonexistent sources, and present false information as truth.

No inference engine does not guess. It does not predict. It does not hallucinate. Every response is retrieved or assembled from a known source: a curated knowledge base, a Wikipedia article, a template, or a deterministic handler. If the system does not have the information, it tells you it does not know.

This makes no inference fundamentally safer for applications where accuracy matters. No fake court cases. No invented citations. No confident lies. Just real, factual information from traceable sources.

▶ What it does

What it does

Module	What it does
Intent detection	Regex-based routing that classifies queries into math, roleplay, instruction, factual, and more.
Knowledge base	Curated JSON files with question-answer pairs across science, history, technology, coding, and everyday life.
Wikipedia	On-the-fly Wikipedia retrieval with an NLG pipeline that strips formatting artifacts.
Template engine	Context-aware templates with topic extraction from conversation history.
Pattern matcher	Social and emotional response patterns for greetings, feelings, and chit-chat.
Fact memory	Extracts and stores user-stated facts (“I like pizza”) for later recall (“What do I like?”).
Math solver	Multi-strategy solver for arithmetic, distance-rate-time, percentages, and probability.
Roleplay engine	Character personas with in-character responses. Einstein, pirate, Shakespeare, and more.
NLG pipeline	Five-pass natural language generation: clean, parse, discourse, realize, fluency.
External APIs	Free public API integrations for weather, time, dictionary, exchange rates, and jokes.
Poem generator	Template-based poem generation. Results range from charming to baffling.
False premise detection	Identifies pseudoscience and non-existent concepts before they reach the knowledge base.

▶ Quick start

Quick start

git clone https://github.com/TheShovel/no-inference
cd no-inference/src && python3 cos_tui.py

Works with the Python standard library. No pip install required for the TUI.

▶ Benchmarks

Benchmarks

NLG Quality (deterministic)

Metric	Score
Fact Preservation	99.1%
Coherence	90.0%
Numerical Precision	97.4%
Temperature Variety	70.0%
Overall	97.4%

LLM Judge (gemma4:31b-cloud, 10 cases, 2 rounds per case)

Metric	Score
Naturalness	8.6 / 10
Informativeness	8.3 / 10
Coherence	9.5 / 10
Correctness	10.0 / 10
Conciseness	9.9 / 10
Overall	9.2 / 10

We prefer a robot that tells the truth awkwardly over one that lies fluently.

▶ CONTRIBUTING.TXT - Notepad

========================================
   CONTRIBUTING.TXT
   How To Help The Project
========================================

=== OVERVIEW ===

This project lives on community effort.

The code is simple. The data is what makes it smart.

We welcome pull requests -- especially ones that add data.

=== WAYS TO HELP ===

> Add knowledge base entries in JSON format

> Write conversation patterns for things the system

does not handle well

> Improve the NLG pipeline to produce more natural

sentences

> Generate more Wikipedia-backed entries

> Fix bugs and write tests

=== HOW TO SUBMIT ===

1. Fork the repository

2. Create a feature branch

3. Make your changes

4. Submit a pull request

5. Wait for review

6. Celebrate!

=== TIPS ===

If you want to add data, look at the existing JSON

files in the knowledge base directory for format

examples. If you want to improve the engine, start

with the intent classifier -- that is where most

of the low-hanging fruit is.

[EOF]

▶ research.html

This project stands on the shoulders of giants! Here be the ancient texts and sacred scrolls that inspired this here engine.

THE READING LIST
1966	ELIZA (Weizenbaum) -- The O.G. chatbot. Pattern matching before it was cool.
1982	Chat-80 (Warren & Pereira) -- Symbolic NL query processing. Old school cool.
1995	Cyc (Lenat) -- When you need 1,000,000 rules to be "smart".
2003	AIML/ALICE (Wallace) -- Pattern-template chatbots for the masses.
2003	SPUD (Stone et al.) -- Because natural language needs a pipeline.
2017	DrQA (Chen et al.) -- Wikipedia as a knowledge source. Read all about it.
2023	MT-Bench (Zheng et al.) -- Multi-turn eval for the modern age.

Last updated: 29-JUL-2026