Beyond Chunk and Pray

Building trustworthy RAG with geometric memory.

Most RAG is "chunk and pray" — split the docs, embed them, retrieve the top-k by similarity, and pray the model grounds its answer instead of hallucinating. This is the open alternative. Retrieval is triple-mediated: questions are answered through a verified knowledge graph with provenance, not by a vector lookup the generator is trusted to use well. The dense index is distrusted — optional, off by default, flagged when used. Every answer is verifiable and cited, numbers are byte-exact, and the system abstains rather than guess when it can't ground a claim.

The baseline runs with no license and no GMS. For production-grade grounding — geometric retrieval, Exact Numerical Memory, contradiction detection, signed answers — the optional GMS backend (knowlytix) snaps in via a lazy seam. Clone it, run the baseline, see exactly where GMS lifts grounding and recall.

Part of the "Beyond … and Pray" series: governed agents · trustworthy RAG · test & validate · LLMs from scratch

Status: the open groundloop baseline (naive chunk→embed→top-k RAG, the deliberate "before") is implemented — numpy-only TF-IDF retrieval + a stuff-and-pray answerer that never abstains. On the Northwind cohort it scores ~20% answer accuracy and 0% abstention: the gap GEODE/GMS closes. Run python demos/naive_rag.py.

What's inside

• Grounded retrieval — answers mediated through a verified knowledge graph
• Provenance + citations — every claim traces to its source
• Byte-exact numbers — Exact Numerical Memory, not "close enough"
• Abstention — says "I can't ground that" instead of guessing
• Distrusted dense index — optional, flagged when used
• GMS-optional — baseline runs free; geometric guarantees via knowlytix

Install

pip install groundloop                # naive RAG baseline (the "before")
pip install "groundloop[ml]"          # + embedders / open-weight models
pip install "groundloop[gms]"         # + the licensed GMS backend (knowlytix)

Quickstart

groundloop is the deliberately naive "chunk and pray" baseline — retrieve top-k, answer, and (tellingly) never abstain. That failure mode is the point:

from groundloop import NaiveRAG

rag = NaiveRAG("Q3 cloud revenue was 120 million dollars. Headcount grew to 340 employees.")
ans = rag.answer("What was Q3 cloud revenue?")
print(f"{ans.decision}: {ans.answer!r}")   # answer: '3'  <- grabs the wrong span, still "answers"

The baseline confidently returns a wrong span and never says "I don't know" — the before the book improves on. The grounded, abstaining GMS approach (verified knowledge graph, provenance, byte-exact numbers) is the Pro tier; see the chapter notebooks in notebooks/.

The GMS upgrade (open-core)

groundloop runs fully without a license. GEODE-RAG — geometric retrieval, Exact Numerical Memory, provenance — requires the licensed knowlytix package, imported lazily:

import groundloop.gms as gms
gms.available()   # True if the licensed backend is installed

Many notebooks — the *_project build variants and the DoE/embedding/baseline appendices — exercise GEODE-RAG and need the backend. It's a one-time setup: get a free developer license, install knowlytix, and every GMS notebook runs.

The production-grade, GMS-native edition is the Beyond Chunk and Pray, Pro Edition — see knowlytix.ai.

License

Apache-2.0. © 2026 Knowlytix.