Skip to main content

Research

How we build, evaluate, and improve Zeever.ca. We publish our methodology, results, and what we learned along the way.

September 8, 2026

Our Vector Index Had Never Been Used Once

For months every search read all 58,000 rows. The answers were correct the whole time, which is why nobody noticed, and it meant a hosted vector database had been compared against a database with no working index.

September 8, 2026

My GPU Benchmark Was Measuring a Cold Model

We published a comparison of two GPUs running local language models. One finding was wrong, because the first request after a model loads under-reports prompt processing by about 25 percent and nothing warns you.

September 3, 2026

Canada Now Has Two Data Centre Rulebooks. Only One Can Say No.

In three weeks Ontario proposed a screening playbook and Ottawa launched a national framework. They agree on almost everything and can do almost nothing to each other. One gates access to the grid. The other is a set of expectations with no metric, no auditor and no consequence.

August 31, 2026

My Toronto.ca Corpus Contained Other People’s Websites

Toronto.ca redirects some of its own pages to other organisations, and my crawler stored what came back under the toronto.ca address. Fixing that deleted 57 real pages and uncovered 13,188 ghost vectors in the search index. Cleaning up the corpus then moved retrieval accuracy eight points, which is more than seventeen models moved between them.

August 30, 2026

I Benchmarked 17 Models. The Problem Was My Benchmark.

Seventeen models scored within nine points of each other on my Toronto question set. Then I checked the benchmark itself and watched retrieval accuracy fall from 89 percent to 68 to 56. Here is what the measurement was hiding, including an entire section of Toronto.ca I had never crawled.

August 23, 2026

Toronto Is 315 MW: What CBRE’s Data Changed in Our Data-Centre Research

CBRE now reports Toronto and Montreal on the same basis as the U.S. markets. We checked our 11-part series against it, corrected six things, and learned that most data-centre arguments are really arguments about definitions.

August 12, 2026

AI Isn’t the Next 2008. But Its Financial System Is Starting to Rhyme.

Data-centre securitization went from $4 billion in 2020 to roughly $61 billion today, and SEC staff have now said qualifying versions are not legally asset-backed securities. This is not subprime. The risk is opacity.

July 31, 2026

Turbopuffer vs pgvector: Equal Quality, and a Speed Claim We Got Wrong

We benchmarked Turbopuffer against our pgvector setup on 100 Toronto questions. Answer quality was a tie, and still is. The 20x speed advantage was not: our pgvector index had never been used, so the comparison measured an unindexed table.

July 28, 2026

Does Proprietary Beat Open? Testing Cohere Command on City RAG

We put Cohere's three Command models against our open Qwen2.5-7B baseline on 100 Toronto questions, scored by a neutral judge. The open model held its ground.

Series · 11 chapters · July 23, 2026

Canada's Data Centre Race

An 11-part research series on the physical foundations of the AI buildout in Canada: data centres, power, fibre, water, sovereignty, jobs, capital, and the alternatives. Grounded in primary sources, with every figure cited.

May 2, 2026

Mapping Canadian AI Compute: Why We Built the Zeever Compute Index

A verified inventory of 39 Canadian GPU providers, normalized on H100 USD/GPU·hr against a $7.50 ceiling, with a sovereignty taxonomy that makes the procurement trade-off legible.

April 23, 2026

The Shift to Agent-First AI: What Model Deprecations Tell Us

AI inference platforms are deprecating chat-first models and replacing them with agent-first MoE architectures. What this shift means for production AI.

April 17, 2026

Running Ollama on an Old Dell XPS With an NVIDIA 3070

We tested local inference with Ollama on a Dell XPS with an RTX 3070 against Fireworks, Together.ai, and OVHcloud. The old desktop held its own.

April 3, 2026

Switching Inference Providers: A 24-Hour Latency Test

We ran a 24-hour latency test across Fireworks, Together, and OVHcloud. Together.ai was 1.7x faster. Here is why we switched.

April 2, 2026

The Missing Layer: AI Inference in Canada

Canada has GPU hosting but no easy way to test and prototype open-source models. That gap is pushing Canadian companies toward US-hosted black boxes.

March 31, 2026

100 Questions Across Toronto.ca: Building a Comprehensive Benchmark

We expanded our evaluation from 12 building permit questions to 100 prompts across 16 categories covering all of Toronto.ca.

March 30, 2026

Scaling a RAG Pipeline from 174 Pages to 35,000 Documents

Our database hit 25GB, the server crashed from OOM, and the web admin went dark during crawls. Here is how we fixed all three.

March 28, 2026

How We Built an LLM-as-Judge to Replace Keyword Scoring

Keyword matching scored our best answer at 0.00. The LLM judge scored it 1.00. Here is how we built a semantic evaluation system.

March 28, 2026

Comparing 7 Open-Source Models for RAG on City Data

We tested 7 open-source LLMs on Toronto city services questions. The cheapest model won.

March 27, 2026

Fixing Vector Search: Probes, Chunking, and Classification

Three retrieval fixes that improved citation accuracy by 35% and fixed queries that returned zero results.

March 27, 2026

Vector RAG vs GraphRAG on Toronto City Data

We compared plain vector retrieval against graph-enhanced retrieval. Graph mode helped on some queries and hurt on others.

Methodology notes

  • All evaluations use temperature=0 for deterministic output.
  • The judge model (GPT-oss 120B via Fireworks.ai) is the same model used for answer generation. We acknowledge the potential for self-evaluation bias.
  • The 100-prompt benchmark suite covers 16 categories across all sections of Toronto.ca.
  • Groundedness scoring evaluates whether claims are supported by retrieved evidence. It does not verify that the evidence itself is current on Toronto.ca.

Code and data

The evaluation framework, benchmark prompts, scoring code, and model comparison scripts are available on request. Contact us at infozeever [dot] ca.