Reports

The Reports page lists every completed evaluation and comparison of the project, newest first. It’s the audit trail of how a model was chosen: which models were tested, on which questions, and which one won. Each row shows the date, the type (Single Evaluation or Pairwise Matrix for comparisons), the models evaluated and the winning model.

View a report

Click View Audit to open a summary:

for an evaluation: hit rate, recall, precision, MRR and nDCG at the top k used, and the number of questions,
for a comparison: each model’s recall, MRR and nDCG.

For the full question-by-question detail, open the run in Evaluation or Comparison, or export the PDF report.

Export a PDF report

Open the export dialog

On the Reports page, click Export on the row of the evaluation or comparison. You can also click Export PDF from the View Audit summary.

Choose Full PDF Report

Truvec loads every result of the run and builds the report in your browser. It takes a few seconds, longer for large golden datasets.

Share the file

The PDF downloads as truvec-<evaluation|comparison>-<name>-<date>.pdf. It is self-contained: share it with stakeholders or attach it to a design review or security sign-off.

The report contains every detail of the run, not only the headline numbers.

Evaluation report

Section	What it contains
Summary	Project, golden dataset, number of questions, top k, number of documents, chunking settings, run dates and duration, and who exported the report. A plain-language verdict with a quality rating, and the key figures: hit rate, recall, precision, MRR, nDCG, F1, latency, costs, vector storage and score gap.
Retrieval quality	The hit rate curve at every cut-off, the rank of the first relevant chunk, and the same numbers as a table.
Score separation & latency	Mean score of the best relevant chunk versus the best distractor, query embedding time, and search time (mean, p50, p95).
Cost & index footprint	Price per 1M tokens, cost to index the corpus, cost per 1,000 queries, cost billed by the run, token counts, number of vectors, dimensions and storage.
Question-level results	Every question with the rank of its first relevant chunk, reciprocal rank, precision, recall, nDCG, top score and search time.
Missed questions	Each question with no expected chunk in the top k, its expected chunks, and the chunks retrieved instead.
Appendix A: retrieval details	For every question, the expected chunks and the full ranked list of retrieved chunks, with file, similarity score, relevance and an excerpt.
Appendix B: methodology	The run configuration, the models (provider, dimensions, price), metric definitions and notes, for audit and reproducibility.

Comparison report

A comparison report covers the same ground for every model, and adds:

Section	What it contains
Recommendation	The winning model, its lead over the runner-up, the models that are best on quality, index cost, query cost, speed and storage, and a leaderboard.
Leaderboard	All quality metrics for every model, with the best value of each column highlighted. Failed models are listed with their error.
Charts	Quality metrics side by side, hit rate curves, first-relevant-rank distributions, quality versus index cost, cost per 1,000 queries, latency and storage.
Head-to-head	For each pair of models, the number of questions where one ranked a relevant chunk higher than the other, and how often the models agree: all hit, all missed, only the winner hit, or the winner missed while others hit.
Question matrix	The rank of the first relevant chunk for every question and every model.
Appendix	For every question, the expected chunks and each model’s ranked retrieval, including the top distractor when a model’s first result isn’t relevant.

If documents were added or removed after the run, the report shows a warning: its numbers describe the corpus as it was when the run started. Run the evaluation again to measure the current corpus.

Costs appear for every model with a known price. See Model pricing.

Limitations

The report uses standard PDF fonts, which cover English and Western European languages. Characters outside them, such as emoji or Chinese, Japanese and Korean text, are replaced by ?.
Evaluations run before detailed metrics were recorded don’t include the curves, latency and cost sections. Run them again to get a complete report.

Shareable links

Sharing a report as a secure link isn’t available yet. Export the PDF, or retrieve results with the API.

Get started

Core concepts

Guides

Reading your results

Resources

View a report

Export a PDF report

Evaluation report

Comparison report

Limitations

Shareable links

​View a report

​Export a PDF report

​Evaluation report

​Comparison report

​Limitations

​Shareable links

View a report

Export a PDF report

Evaluation report

Comparison report

Limitations

Shareable links