Open source · Vectorless on-device RAG

Turing
Tree

No vectors. No chunking. No blind trust.

A vectorless, reasoning-based RAG that runs entirely on your machine. Drop in your documents and ask — every answer comes back grounded in your sources, with a confidence score that tells you when to trust it.

Get started View on GitHub

100%

on-device

silently-wrong answers

API keys · no cloud

document formats in

The benchmark story

Same recall as a vector database —
but it knows when it's wrong.

In a controlled run on web-crawled Wikipedia science articles, Turing Tree retrieved with the same embeddings as a classic vector DB — so it found exactly what the vector DB found. Then it flagged every off-topic question the vector DB answered silently and wrongly.

1.00

accuracy@5 — identical retrieval to a vector DB (0.00 cross-contamination)

100%

of off-topic questions abstained — the vector DB flagged 0%

+79.5

confidence separation (of 100) between trustworthy and confused retrievals

198×

faster re-index of unchanged content via the content-hash cache

Vector DB baseline

Returns a confident-looking top-k for every question — including ones the corpus can't answer. Off-topic queries come back silently wrong.

Turing Tree

Measures the shape of each retrieval and abstains on 100% of off-topic queries — while never abstaining on a real in-corpus one.

Measured: 6 docs / 90 chunks, local Ollama, top-k = 5. Reproduce with python scripts/benchmark_rag.py — full method & caveats in docs/benchmark.md.

Why it's different

A retriever that reasons, scores, and abstains.

Three choices set Turing Tree apart from a normal RAG chatbot.

Vectorless retrieval

No vector database. Each document becomes a hierarchical reasoning tree the model navigates by thinking — not nearest-neighbour lookup.

Confidence you can act on

Every answer carries a 0–100 score and a verdict — Answer, Review, or Abstain — with the drivers behind it. Off-topic questions are flagged, not faked.

100% on-device

Local Qwen models via Ollama. Nothing you upload or ask ever leaves your computer. One command to set up, one to run — no API keys, no GPU.

Secure & trustworthy by construction

Privacy isn't a setting — it's the architecture.

Because everything runs locally and every answer is scored, both safety and reliability are properties of the design, not promises.

How it works

From your documents to a grounded, scored answer.

Your documents

Upload files or a whole folder — PDF, DOCX, Markdown, HTML, or text.

Reasoning tree

Each document is turned into a table-of-contents tree the local model reasons over.

Grounded answer

You get a cited answer and a confidence verdict — or an honest abstain.

Photosynthesis converts light into chemical energy stored as glucose.

92 · ANSWER photosynthesis.md

Get started

Two commands. No keys. No cloud.

Clone, run setup once, then launch. It builds the UI the first time and opens your browser at 127.0.0.1:8765.

View the source

# 1 — clone
git clone https://github.com/1ssb/TuringTree RagIndex
cd RagIndex

# 2 — set up everything (venv, deps, models, Ollama)
$ bash scripts/setup.sh
#   Windows: powershell -File scripts\setup.ps1

# 3 — run (opens http://127.0.0.1:8765)
$ ./run.sh
#   Windows: run.bat

Explore the source on GitHub.

Turing Tree is open source. Clone it, run it locally, and read the full method and benchmarks.

Open on GitHub