How it works
Behind a Medetatio answer.
Language models, checking, and verification are parts of the same broader workflow. Here is the initial overview, with room for the technical details to follow.
In plain language
Checking is part of the system.
Medetatio uses language models as part of a broader machine-learning workflow. That workflow also includes cross-encoder components and additional checkers, with the aim of reducing hallucinations and low-quality responses.
The full explanation will describe what is checked, how those checks affect a response, and where the system can still fail.
- Language-model components
- Used within the response workflow. Their roles and how they interact with the checking systems will be documented.
- Cross-encoder components
- Included in the machine-learning workflow. Their exact role, inputs, scores, and effect on responses will be documented.
- Checkers and verification
- Additional systems intended to identify hallucinations and low-quality output. The specific checks and decision rules will be documented.
A closer look
Technical outline
The following topics are reserved for the technical description. The actual processing sequence, methods, and results will be supplied by the development team.
Components and processing flow
Model and checker roles, their inputs and outputs, and the actual processing sequence.
Technical description to be added.
Cross-encoder components
What is scored, how scores are used, and model details and versioning where shareable.
Technical description to be added.
Checking and verification methods
What each checker tests, the evidence it uses, decision criteria, and known blind spots.
Technical description to be added.
When a check fails
How the system handles weak evidence, conflicting checks, or a response that does not meet its quality criteria.
Technical description to be added.
Sources, evidence, and citations
How references are selected and connected to claims, what materials are included, and what a citation establishes.
Technical description to be added.
Evaluation and results
Test sets, quality measures, comparison methods, error rates, and reproducible examples where available.
Evaluation methods and results to be added.
Reading the results carefully
Checks reduce risk. Errors can still happen.
The checking workflow is intended to improve response quality. It does not guarantee that every claim, reference, or interpretation is correct.
- Review important claims and read the cited passages in context.
- A citation or model score should not be treated as proof on its own.
- Measured performance claims need their test methods, conditions, and limitations alongside them. Those results are awaiting documentation.
Questions about the system?
For technical questions or conversations about using Medetatio, get in touch.