Built for Rigorous Qualitative Research
QUAILS is a free, open-source desktop application that helps researchers apply AI assistance to qualitative text analysis that runs systematically, transparently, and entirely on their own computer.
QUAILS is a free, open-source desktop application that helps researchers apply AI assistance to qualitative text analysis that runs systematically, transparently, and entirely on their own computer.
QUAILS (Qualitative AI Labeling Support) is an app that runs on your own computer for automating thematic text analysis in qualitative research. It uses large language models (LLMs) for inference that can be run either locally on your own computer for free (via a free software called Ollama), or through cloud providers like OpenAI, Anthropic, and Google Gemini.
QUAILS provides a visual flow editor for designing multi-round qualitative coding analysis workflows and enforces human review checkpoints between analysis rounds, keeping researchers in control of the analytical process at every stage.
QUAILS was designed for qualitative researchers who work with large text-based datasets like piles of interview transcripts, open-ended survey responses, field notes, policy documents, social media posts, or any other corpus where manual review at scale is a bottleneck and is challenging to conduct and time consuming, even with a large team. It helps researchers get those large-scale analyses done that would otherwise be impossible!
Qualitative researchers increasingly face data corpora that are too large for traditional manual coding, even with large teams at their disposal. This can be hundreds of transcripts, thousands of survey responses, or entire policy archives. QUAILS does not replace the human in the qualitative research process or the ability of humans to interpret meaning. Instead, it helps researchers tackle large datasets that would never see analysis in the first place. For automated qualitative research tools, the answer is not to abandon rigor and procedure by giving it to an AI and simply trusting its summaries. The solution is instead to apply AI assistance in a way that preserves rigor, replicability, and human centering to prioritize researchers in the process and conduct transparent science.
QUAILS is not a replacement for qualitative expertise. It is a tool for extending that expertise to datasets that would otherwise be impractical to analyze rigorously by hand.
Every design decision in QUAILS reflects this philosophy. The alignment testing workflow forces calibration against the researcher's own judgement before full analysis begins. Human Review nodes are automatically inserted between pipeline steps. Confidence thresholds let the researcher decide when the AI's certainty is high enough to trust its output.
The researcher writes the prompts, sets the thresholds, labels the calibration sample, inspects the output, and decides whether to proceed. The AI does the repetitive work at scale, not the interpretive work that requires domain knowledge and theoretical grounding.
The Alignment tab enforces a calibration step before full analysis. You human-label a random sample, compare these results against the AI with Cohen's Kappa or percent alignment, refine prompts, and repeat until agreement is acceptable.
Every label carries a 0–100% confidence score derived from actual system token log-probabilities, or the statistical probabilities that a specific word will be chosen. Set your own thresholds to flag uncertain decisions for your own review.
Every analysis output includes the label, confidence score, and AI reasoning for every unit. It doesn't just provide the final decision, but also an explanation of why the system made it. That transparency runs through the entire data flow pipeline: every prompt, model setting, human-review decision, and pipeline change is recorded and traceable from raw input to final code. These are the receipts behind explainable AI. We designed a system you can inspect, defend, and reproduce, and not just one you're asked to simply trust.
Human review isn't just a cliche with QUAILS. It's the default the analysis pipeline is built around. At every step, the researcher can accept, edit, or reject any AI decision, have items flagged for closer reading, or send a round back for re-analysis with revised prompts. Nothing becomes part of the final dataset without a human person actively approving it: the AI proposes, the researcher decides.
Paste a dataset into a chatbot and you can't be sure every row gets a response. General-purpose LLMs summarize, skip, or merge units when a conversation runs long (or even just a few tokens long), with no log to check against. QUAILS systematically processes every unit you define, be it a sentence, a response in a chain of dialogue, paragraphs, or a whole document. Each unit of analysis (whatever you choose) is its own isolated interaction with the AI and the system accounts for each unit individually in the results table that can be traced back to the source document. This guarantee of complete, systematic coverage isn't something a frontier chatbot interface can offer, because it isn't built to run a procedure: it's only built to hold a conversation.
The technical choices inside QUAILS are not arbitrary: each one serves the goal of producing analysis results that are consistent, defensible, and reproducible.
The system defaults the temperature parameter to 0, meaning the LLM always picks the highest-probability token. Combined with a fixed random seed parameter, this makes analysis fully deterministic (i.e., it will be the same every single time the same analysis is run). Running the same pipeline on the same corpus will return identical results, which is critical for replicable science. It also makes it so another researcher can run the exact same QUAILS configuration on the data and produce the same results.
Confidence scores are derived from actual token log-probabilities (i.e., the statistics on what token/label is selected). Most systems simply ask the AI "how confident are you?", which produces unreliable self-reports that have no way of verifying the output. The formula
e^logprob × 100 converts the model's actual output probability into a 0–100% score for easy human review.
Reasoning-capable models like DeepSeek-R1, Qwen3, and QwQ emit internal "thinking" tokens before their final answer. QUAILS disables this at the API level and strips residual thinking tokens from all responses. This ensures logprob scores reflect only the answer tokens, not the model's internal deliberation, and keeps responses grounded purely in the instructions you write.
Each text unit is analyzed in a fresh, independent call. No conversation history from a previous unit influences the next one, keeping the analysis engine clear and not dependent on prior information it receives, which could corrupt the computational results from the AI if this was present. The system truly evaluates one unit of text at a time, be it a sentence, paragraph, or whole document (you choose!) The only cross-unit information that flows between units is the explicit context window setting, and only when you configure it above zero.
QUAILS runs entirely on your own computer and has no connection to its developers. The application cannot call home. Your documents, prompts, labels, and results stay on your machine.
All AI inference runs on your local hardware. No data ever leaves your machine. No internet connection is required for analysis once a model is downloaded. Your documents, results, API configurations, and API keys remain entirely private on your computer. This is the recommended mode for any data involving human research participants.
When you configure and use OpenAI, Anthropic, Google Gemini, or Azure OpenAI, your document text is sent to that provider's servers for processing. This is subject to each provider's data retention and usage policies. Before using cloud providers with research data, consider your obligations under IRB requirements, HIPAA, FERPA, GDPR, and any other applicable regulations. You are solely liable for any data sharing that occurs through third-party providers configured in this application.
Anonymous crash reports and user suggestions can be submitted to the developer, but only with your explicit, per-event opt-in. These reports contain only the error message, OS type, browser version, and app version. No project data, document text, prompts, labels, API keys, or personal information is ever included.
QUAILS was developed with the support of U.S. federal research infrastructure and AI resources made available to the research community.
QUAILS was designed and built out of the TRAILblazer Lab, a research group at the University of Illinois Chicago investigating the crossroads of AI- and simulation-based technologies for teaching and learning.
QUAILS is free, open-source, and runs entirely on your own machine. Download it and run your first analysis today.