🌙 LUAR: Local Utility for Automated Reviews

Label a spreadsheet of text with your own questions, on your own computer.

You drop a spreadsheet, write what you want to know about each row ("What is this review about?", "Does the customer ask for a refund?"), and LUAR gives you back a copy of the spreadsheet with the answers filled in, plus a short report. No cloud, no account, no API costs: your data never leaves your machine.

Leia em português

Contents

1. What LUAR does
2. Before you start
3. Installing
4. Your first run, step by step
5. Writing good questions
6. Understanding the results
7. Checking whether you can trust the answers
8. How accurate is Laya
9. Using the command line
10. Troubleshooting and FAQ
11. Related projects, contributing and license

1. What LUAR does

Imagine a spreadsheet of customer reviews:

You ask two questions: "What is this review mainly about?" (delivery, product, support or price) and "Does the customer ask for their money back?" (yes or no). LUAR returns a copy of the file with new columns:

_{Real output of the Laya engine on the bundled example. All four answers are right, but the model was unsure about the topic (below the 0.7 threshold), so both rows are marked for a quick human check.}

• Each answer comes with a confidence from 0 to 1.
• Rows where the model was unsure are marked needs_review, so you know where a person should look.
• A summary report counts the answers and lists the rows to review.
• Your original file is never changed.

It is useful for survey answers, product reviews, support messages, feedback forms, notes: any column of text you would otherwise read and tag by hand.

2. Before you start

You need:

• Python 3.10 or newer. Check by opening a terminal (on Windows: PowerShell or Command Prompt) and typing python --version. If it is missing, install it from python.org and tick "Add python.exe to PATH" during setup.
• A spreadsheet in .xlsx or .csv format, with the text in one or more columns.
• Disk space and a one-time download of about 2.5 GB (libraries and Laya's two models). After that, everything works offline.

The "brain" that reads your text is Laya, an open decision model built for this kind of question: instead of writing text, it gives a probability to each option. It runs on your computer, without a GPU, at about 1 second per row.

Laya comes ready to use: it does not need training, and it does not learn from your spreadsheets. What improves the answers is writing good questions (section 5) and measuring accuracy on a sample (section 7).

3. Installing

Open a terminal and run:

pip install "luar[all]"

Installs LUAR with Laya and the web interface. (pip install luar, without [all], installs only the command line, still with Laya.)

To have LUAR ready to use without internet, download the Laya models right after installing:

luar download

It fetches the multilingual and English models (about 1.5 GB, only once). After that, LUAR makes no outside connection: neither the interface nor the model.

On Linux, install the CPU version of PyTorch first, or pip will download a multi-GB GPU build: pip install torch --index-url https://download.pytorch.org/whl/cpu

4. Your first run, step by step

Step 1: open LUAR

luar ui

Your browser opens on http://127.0.0.1:7860. This page is served by your own computer and is not reachable from the internet. Keep the terminal open while you use LUAR; close it (or press Ctrl+C) to stop.

If luar is not recognized, or Windows blocks it, use python -m luar ui instead.

Step 2: load your spreadsheet

Drag your file onto the CSV or XLSX box (or click it to choose the file).

• LUAR shows how many rows and columns it found and a preview of the first rows.
• CSV files exported by Excel in other languages (with ; as separator, or accented characters) are detected automatically.
• For Excel files, the first sheet is used.
• Under Column(s) the model should read, tick the column with the text. LUAR pre-selects the column with the longest text. If you tick several (for example title and body), they are read together.

Just want to try it? If you installed from the repository (see section 11), pick a bundled example in …or try an example, or click Sample CSV at the top of the page to download a test spreadsheet with 20 made-up customer messages, and Sample JSON for the questions that go with it (load them in Load questions (.json), step 3). The README button next to them opens this manual. To load both at once, pick Amostra de teste (Sample CSV) in the examples menu. Otherwise, download reviews.csv and reviews_questions.json and load them as described in steps 2 and 3.

Step 3: write your questions

Each row of the questions table is one question:

Click a cell to edit it, and use the table's row controls to add or remove rows.

Tip: questions can be saved and reused. Save questions as .json downloads the table; Load questions (.json) brings it back. Writing the questions once in a .json file is often the easiest way to manage many of them (format in section 5).

Step 4: choose the settings and run

• Confidence threshold (default 0.7): any answer below this is marked needs_review. Raise it to review more rows, lower it to review fewer.
• Laya model variant: leave it on auto. It picks the English model for English files and the multilingual model for everything else.

Click Run. If you did not run luar download when installing, the first run downloads Laya's model, which takes a few minutes; later runs start in seconds.

Step 5: download the results

Under Result you get two files to download:

• <your file>_luar.csv or .xlsx: your spreadsheet with the new columns.
• <your file>_luar_summary.md: the report (it is also shown on the page, in Summary). .md is plain text, so it opens in any text editor.

The Table tab shows the result spreadsheet right on the page.

5. Writing good questions

The quality of the answers depends mostly on how the questions are written.

The three question types

⚠️ score is experimental. It was the least reliable type in our tests: on 300 real product reviews, Laya matched the customer's own star rating (1 to 5) only 26% of the time, although 60% of its answers were within one star. On an urgency scale (low / medium / high) for support messages it got 3 of 12 right, and neither describing each level, nor reversing the order, nor switching to choice fixed it: some scales are simply beyond the model. Prefer choice or noul when you can (for example "Does the message report lost money or a health risk?" instead of an urgency level), describe each level if you use score, and always measure with expected_ columns (section 7). Because its confidence is almost always low, a score question does not mark rows for review.

Tips

1. Describe the options. price: cost, value for money, charges works much better than just price. The description tells the model what belongs in each option.
2. Make the options cover everything and not overlap. If some rows may fit none of them, add an other option. If two options mean almost the same thing, merge them.
3. Ask one thing per question. Instead of "Is the customer angry and asking for a refund?", make two noul questions.
4. Be specific. "Does the customer explicitly ask for their money back?" gets better answers than "Refund?".
5. Keep to about 20 options. For longer lists, split into a broad question and a detailed one.
6. Option order can change some answers. If results look biased towards one option, try reordering and compare (section 7 shows how to measure).
7. Write in the language of your data. Questions and options can be in any language the engine supports.
8. Be careful with "middle" options such as neutral. On real reviews, Laya got positive and negative right 82–87% of the time but almost never chose neutral (3 of 60). If you need a middle option, describe it well and measure it with expected_ columns (section 7); two yes/no questions ("Is it positive?", "Is it negative?") can be an alternative worth testing.

Questions file format

A questions file is a list in JSON format. This is the file behind the example above:

[
  {
    "id": "topic",
    "type": "choice",
    "question": "What is this review mainly about?",
    "options": {
      "delivery": "shipping, delays, courier, package condition",
      "product": "quality, defects, how the item works",
      "support": "customer service, returns, how the store handled a request",
      "price": "cost, value for money, charges and billing"
    }
  },
  {
    "id": "sentiment",
    "type": "choice",
    "question": "What is the overall sentiment of the review?",
    "options": ["positive", "neutral", "negative"]
  },
  {
    "id": "wants_refund",
    "type": "noul",
    "question": "Does the customer explicitly ask for their money back?"
  }
]

options can be a simple list (["positive", "neutral", "negative"]) or a list with descriptions ({"delivery": "shipping, delays", …}). For score, list the levels from lowest to highest.

6. Understanding the results

New columns in your spreadsheet

For each question (here, a question with id topic):

And two columns for the whole row:

Rows with no text are not sent to the engine; they are marked needs_review with the reason empty text.

The summary report

The _summary.md file contains:

• The run: date, engine, number of rows, columns read, threshold, time taken.
• One section per question: how many rows got each answer (count and %), the average confidence, and how many answers fell below the threshold.
• Accuracy, if your file has answer-key columns (next section).
• Rows to review: row number (as in Excel, counting the header as row 1), which questions were uncertain, and the start of the text. Up to 50 rows are listed; filter needs_review in the result file for the rest.

How much to trust the confidence

The confidence is a real probability computed by the engine, not a number the model makes up. It is a good guide to where mistakes are likely, but how good depends on your data. In our tests, Laya was right 98% of the time when its confidence was 0.7 or more on news topics, but only 72–75% on customer reviews. A model can be confidently wrong, which is why the next section matters.

7. Checking whether you can trust the answers

Before relying on LUAR for a large file, measure it on a sample whose answers you already know:

1. Pick 20 to 50 rows and answer the questions yourself.
2. Add one column per question named expected_ + the question id. For the question topic, the column is expected_topic. Fill in your answers.
   • For noul questions, yes/no, sim/não, true/false and 1/0 all work.
   • Rows left blank in the expected_ column are simply not counted.
3. Run LUAR on that file. The summary now shows, for each question, a line like Accuracy against expected_topic: 12/14 (86%).
4. If the accuracy is too low, improve the questions (section 5): describe the options better, split questions that mix two things, swap score for choice or noul, or try the other Laya model variant. Run again until you are satisfied. Then run your full file.

The expected_ columns are never shown to the model, so they cannot leak the answers.

8. How accurate is Laya

We measured Laya on 300 texts from two public datasets labeled by people, on a laptop without a dedicated GPU (details and script):

• Right when confidence is 0.7 or more: 98% on news, 72–75% on reviews.
• Time per row: about 1 s with one question, 1.3 s with three.

_{Review labels come from the customers' own ratings, which do not always match the text, so they are a hard, noisy test.}

Accuracy depends a lot on the question type and your data: clear choice or yes/no questions do well; scales (score) and "middle" options do poorly. So measure on your sample with expected_ columns (section 7) before trusting a large run.

9. Using the command line

Everything in the web interface can also be done in a terminal, which is handy for repeating the same job or automating it.

luar columns my_file.xlsx

Lists the columns of a file with a sample value from each.

luar download

Downloads Laya's models (multilingual and English) so LUAR works offline afterwards.

luar run my_file.xlsx -q questions.json -c review

Labels the file. The result and the summary are saved next to the original file, as my_file_luar.xlsx and my_file_luar_summary.md. If those names already exist, LUAR adds _2, _3… instead of overwriting.

Use LUAR from Python:

from luar import load_questions, run_file
from luar.backends import make_backend

backend = make_backend("laya")
result = run_file("data.csv", load_questions("questions.json"), ["text"], backend)
print(result.table_path, result.summary_path)

10. Troubleshooting and FAQ

Does my data leave my computer? No. Laya runs on your computer and LUAR makes no outside connection: the interface and the model work offline. The only downloads are the software and the models, when installing (see luar download in section 3).

Do I need to train Laya? Does it learn from my data or my corrections? No and no. Laya comes trained and answers new questions from the text of the question and options alone; it keeps nothing from your spreadsheets or from the corrections you make to the result. To get better answers, rewrite the questions (section 5) and measure accuracy with expected_ columns (section 7).

luar is not recognized, or Windows says access is denied. Use python -m luar ui (or python -m luar run …). Some Windows security settings block small program files that pip creates; going through python avoids that.

"Old .xls files are not supported". Open the file in Excel and save it as .xlsx (or .csv).

Accents look wrong in the result. The result CSV is saved in UTF-8, which Excel reads correctly when you double-click the file. If you import it manually (Data → From Text/CSV), choose UTF-8.

"The file already has column(s) …". You are probably running LUAR on a file that is already a LUAR result. Use the original file, or change the question ids.

"The Laya engine could not be loaded". Laya comes with LUAR, so this means a broken install. Run pip install "laya>=0.3.21" and open LUAR again.

It is slow. Laya takes 30 to 60 seconds to load at the start of each session; after that it handles about one row per second, so a 1,000-row file takes about 20 minutes. To tune your questions, try a small sample first.

Many rows are marked needs_review. Improve the option descriptions (section 5), check accuracy with expected_ columns (section 7), or lower the threshold if accuracy is already good.

Which languages work? Laya's multilingual model handles many languages (it was tested here in Portuguese and English).

11. Related projects, contributing and license

Related projects. Other open tools around Laya, in case one fits you better:

• laya-studio: a full web platform (Docker + PostgreSQL) with batch evaluation, calibration monitoring, a review queue and automations. Pick it if you are building a decision system for a team; pick LUAR if you just want a labeled spreadsheet.
• vgi-laya: Laya as SQL functions inside DuckDB. Pick it if your data already lives in a database and you are comfortable with SQL.
• More in the laya.tools directory.

Installing from the repository (includes the examples and the example picker):

git clone https://github.com/HayateV30/luar.git
cd luar
pip install -e ".[all]"

Contributing. How LUAR works inside, how to run the tests and how to add an engine: CONTRIBUTING.md. Bugs and ideas: issues.

License. MIT for LUAR's code. Laya (the package and the model weights) is a separate project with its own license; check its model card before redistributing it.