Explanation generator
Purpose and deployment context
The Explanation generator converts the technical components of the ranker score into a single short Czech sentence for the candidate. It is used on the position detail after search and helps to fulfil the right to a comprehensible explanation. It does not evaluate the candidate for the employer and must not promise an interview, acceptance or any other recruitment outcome.
Classification under the AI Act
Under Regulation (EU) 2024/1689 (the AI Act), we assess this component in its current deployment as probably outside Annex III, point 4(a). The generator only converts the components of the ranker score into a comprehensible sentence for the candidate; its output is shared exclusively with the candidate and does not serve for the evaluation or selection of candidates on behalf of an employer. The classification shares the assessment of the matching ranker, we keep it under ongoing review and it is subject to a final legal assessment. If, in the future, the generator became part of decision support for the selection of candidates on behalf of an employer, it would be reclassified as high-risk under Annex III, point 4(a), with the full AI Act regime; the provider obligations would apply from 2 August 2026. Even outside this regime, the Operator voluntarily maintains technical documentation of the prompt and guardrails, a model card, logging, post-market monitoring and an incident process; serious incidents would be reported to the competent authorities under the AI Act and Czech implementing legislation.
Inputs and outputs
The input is the position ID, the position title, the first 500 characters of the description, up to the three strongest score components and an optional candidate summary of up to 200 characters. The prompt redacts e-mails, telephone numbers, birth numbers and selected personal identifiers. The output is a single Czech sentence without markdown, bullet points or emoticons; the result is returned with the model, the cache source and the time of generation.
Models used and versions
The explanation is generated by a fast LLM. By default the Claude 3.5 Haiku model (Anthropic), alternatively GPT-4o mini (OpenAI) or a corresponding model via Azure OpenAI as an EU-resident variant. No fine-tuning is recorded. The knowledge cutoff depends on the model provider.
Training data
JobsAI does not train or fine-tune the underlying LLM. The in-context prompt contains system instructions in Czech, rules against promises of a recruitment outcome, protection against prompt injection and a data block with the position title, an excerpt of the description and the score components. The full wording of the prompt is maintained internally; the public card describes the purpose, inputs and security limitations.
Limitations and known risks
The LLM may hallucinate, over-generalise the match, ignore nuances of seniority or repeat problematic text from the advertisement. Validation removes markdown, checks the length and prohibits emoticons, but does not guarantee complete factual correctness. On a timeout, a provider error, quota exhaustion or a score error, the UI degrades without generated text.
Fairness and anti-discrimination
The prompt prohibits speculation about the candidate's identity and must not use protected characteristics under Act No. 198/2009 Coll. The input does not contain the candidate's name and is to use only the score components and the job context. The pre-launch fairness audit of the ranker and the filter is completed according to the internal methodology. Last audit: 2026-05-14, passed.
| Item | Value |
|---|---|
| Last audit | 2026-05-14, passed |
A manual review of a sample of explanations from real data remains for Q1 after launch.
Human oversight
The candidate can ignore the explanation and submit a complaint. An administrator can investigate the audit record of an explanation, temporarily switch off the generation of explanations, reduce the quota or block a specific problematic position. The explanation is supporting text, not an automated decision.
Performance metrics
The numbers of requests are measured by source fresh, cache and degraded, together with the estimated cost and latency in internal operational metrics. Qualitative accuracy and the hallucination rate have not yet been measured; measurement will take place from the Sprint 5 launch window through a manual QA sample and user complaints.
Security and privacy
The processing of personal data is governed by Regulation (EU) 2016/679 (GDPR); for the explanation of a match to a candidate, the legal basis is the performance of a contract under Article 6(1)(b) GDPR, whereas for auditing, transparency and improvement of the service, the legal basis is legitimate interest under Article 6(1)(f) GDPR. Anthropic direct, per the documentation, crosses the EU boundary to the USA and requires SCC/ZDR and PII stripping. OpenAI direct may also mean processing outside Europe depending on the chosen account; Azure OpenAI is the preferred EU-resident variant via deployment. Results are cached for 24 hours and the explanation audit is purged according to the retention job for search logs. Details are in the privacy policy.
Complaints and contact
Send complaints about a misleading or discriminatory explanation to [email protected]. Direct requests for access to or erasure of the related personal data to the privacy contact in the privacy policy.
Card version
Card version 1.0.3, last updated 2026-07-08.