
Inference
Running an already trained model to produce output. It differs from training; its cost depends on the model and processed volume.
#Short, precise definitions to understand product claims, ask better questions and evaluate tools. 99 terms organized by use.


Running an already trained model to produce output. It differs from training; its cost depends on the model and processed volume.
#
A document describing a model, intended uses, limitations and information about training or evaluation. Check which version it covers.
#
A system trained on data to produce an output from an input. One model can be offered through several applications.
#
A model trained to process and generate language. It can write or summarize, but it is not a guaranteed database of facts.
#
A unit of text processed by a model: a word, part of a word or punctuation. Limits and costs are often expressed in tokens.
#
The maximum amount of information a model can consider in one request, including input and output.
#
A response following a defined format, such as a JSON schema. Valid formatting does not prove the values are correct.
#
An attempt to make a model follow instructions inserted into input or documents instead of the authorized task. Treat external documents as untrusted data.
#
Instructions and context given to an AI system. A useful prompt states the intended result, constraints and format.
#
An instruction shaping an assistant’s behavior. It does not replace technical access and data controls.
#
The ability to process or produce several content types, such as text, images, audio and video.
#
A parameter affecting output diversity in some models. A lower value does not guarantee accuracy.
#
Extracting vocals or instruments from an audio mixture. Separated stems may retain residual sound or artefacts.
#
Sound from another target remaining in a separated stem. Evaluate it in isolation and its final usage context.
#
Number of audio samples per second, expressed in hertz. Do not confuse it with compressed-file bitrate.
#
A structured description of expected properties and constraints in a JSON document. Schema validation does not prove factual accuracy.
#
Checking an output’s structure: required fields, types and constraints. It complements meaning checks without replacing them.
#
Converting characters in an image into text. Recognition does not validate reading order, table structure or value accuracy.
#
Reordering retrieved documents by relevance to a query. It does not replace checking the content of the documents.
#
The origin and history of information: document, author, date and transformations. Provenance helps trace the support for a claim.
#
How well a source’s date fits the question. Distinguish publication date, event date and the corpus’s last synchronization.
#
A plausible but false, invented or unsupported answer. A source or citation can also be fabricated.
#
A method linking an answer to identifiable documents. It helps verification, but a source can still be misread.
#
An architecture retrieving passages from a corpus before passing them to a model for its answer. Quality also depends on the documents and retrieval.
#
A reference shown with an answer so a claim can be traced. Open the source and check the passage, date and context.
#
A document-object identifier used to retrieve its reference and metadata. A DOI does not certify content reliability.
#
A paper selected as the starting point for relationship-based discovery or recommendations. Seed choice influences discoveries.
#
A representation of relationships between publications. Identify connection meanings: citation, computed proximity or recommendation.
#
Information describing a document or resource, such as title, date, authors or identifiers. Metadata can be incomplete or incorrect.
#
Scope of documents actually present and usable in a collection. One correct answer does not establish topic coverage.
#
An identifiable document state at a particular date or revision. Two versions can contain different conclusions.
#
Rules determining which documents enter a corpus or review. Their application should be explainable and retained.
#
Signed metadata linking claims about provenance or edits to a media file. Verify what the media appears to depict independently.
#
Text encoded in a document, distinct from the visible page image. Selectable text can still be incomplete or have an incorrect order.
#
The sequence in which headings, paragraphs, columns and notes should be read. An extractor can recognise every word but order them incorrectly.
#
Representing a parsed document in a form suitable for storage or exchange. Lossless serialisation of a parsed document does not correct recognition errors.
#
A text file written alongside a PDF. It makes OCR output easier to inspect. Check page alignment, especially when a page was skipped or already contained text.
#
An XML representation following Text Encoding Initiative conventions. Its elements can identify an article body, sections and references.
#
Matching an extracted reference against bibliographic metadata to supplement or check an identifier, title or date. Review ambiguous matches and shared names.
#
A numerical representation used to compare semantic similarity. Nearby items are not necessarily equivalent or true.
#
A system storing vectors and retrieving nearby items. In RAG it helps select passages but does not verify their accuracy.
#
Additional training on targeted examples to change model behavior. It does not replace an up-to-date document collection.
#
A technical interface through which an application requests an AI service. Costs, quotas and data rules depend on the plan.
#
Software executing a model to produce outputs. The same model can behave differently across engines and configurations.
#
Representing weights with reduced numerical precision. Resource requirements and quality may change; configurations require testing.
#
A format used by some tools to distribute model weights and information. It does not guarantee compatibility with every engine.
#
A service or repository distributing identified models or artefacts. Retain version, provenance and usage conditions.
#
An address providing access to an API function. Compatible endpoints can have different models, quotas and processing conditions.
#
Progressive transmission of output before completion. Receiving an early fragment quickly does not measure complete duration.
#
Number of simultaneously active operations in a measurement or system. Fix it when comparing workloads.
#
Grouping multiple inputs for processing. Batch measurements are not directly equivalent to interactive conversation latency.
#
An initial run requiring setup or loading. Distinguish its delay from an already ready service.
#
Reuse of computation for matching beginnings of requests. Measure requests that benefit separately from requests with fresh prefixes.
#
A restriction on permitted request frequency or volume. Record its window and scope before comparing plans. Compare short-term limits with the permitted volume over the full period. A short trial may succeed while simultaneous users exceed a shared limit; test that case without running unauthorized load.
#
Temporary retention of results or computations for reuse. Check duration, information freshness and data handling. Check example: after correcting the source document, ask for the same information again and check the version used. An unchanged answer may come from a retained copy rather than a fresh reading.
#
Explicit selection of a dependency or model version to limit unintended changes. It does not guarantee reproducibility of every output.
#
An environment limiting program access or actions. Qualify actual boundaries instead of assuming complete isolation.
#
A label such as application/pdf that identifies a format. Compare the declared type with detected contents during ingestion and inspect ambiguous files.
#
A family of PDF profiles designed for long-term preservation. Check the required profile and content fidelity, including the text layer when relevant.
#
A property where repeating an operation has no additional effect compared with executing it once. Useful for preventing duplicates during retries.
#
Granting only permissions needed for a task and for the required duration. An agent should not receive all access by default.
#
Re-execution after failure. Limit attempts, distinguish temporary from permanent errors and prevent duplicate writes.
#
A system combining a model, goals, tools and decision steps to complete a task. Its autonomy should be bounded by permissions.
#
A mechanism through which a model asks an application to run a function such as search or calculation. The application controls execution.
#
A sequence of steps connecting people, data and tools. Reliable automation plans for errors, recovery and approval.
#
A decision or check performed by a person before or after a system action, useful when errors could have significant consequences.
#
A component connecting a tool to another application or source. Check authentication, permissions, read data and possible actions.
#
A workflow used when a component fails or is unsuitable. Check that fallback does not silently change important conditions.
#
Waiting between repeated attempts after failure. It should include a limit and checks against duplicate effects.
#
A description of process states and allowed transitions. It helps distinguish waiting, execution, failure and accepted results.
#
A process that produces output while failing on some elements. Check coverage and warnings: an output file does not validate all its pages.
#
An error whose consequence makes an output unacceptable even if other criteria are satisfied. Define it before the trial, such as a wrong recipient, incorrect amount or ignored permission.
#
Declining to answer or act when information or authorisation is insufficient. Evaluate it on explicitly out-of-scope cases.
#
Observable conditions set before a trial to determine whether a result is usable. Include errors that require rejection.
#
Total trial cost divided by results passing checks, including service, calls, preparation and correction. It differs from generation price.
#
Understanding system behaviour through logs, metrics and traces. Retain useful evidence without unnecessarily logging sensitive data.
#
A stable set of representative cases and criteria used to compare versions of an AI system on a defined task.
#
A standardized evaluation comparing systems under a stated method. A strong public score does not guarantee quality on your data.
#
A decline in behavior that previously worked, sometimes after a change to a model, prompt or data.
#
The time between a request and a usable result. Also measure retrieval, tool execution and human review time.
#
Delay before receiving the first generated textual element in a stream. It represents neither total duration nor response quality.
#
Work processed per unit of time under a defined scope and workload. Distinguish produced from accepted results.
#
A time limit after which an operation is interrupted or reported failed. Recovery should avoid duplicating completed actions.
#
Expected answer or facts used to check a trial. Acceptable wording can vary without changing required facts.
#
Different judgments of the same output. Inspect criteria and evidence rather than hiding disagreement in an average.
#
The share of retrieved documents relevant to the question: relevant hits divided by all retrieved documents. It does not assess generated-answer accuracy.
#
The share of relevant corpus documents that were retrieved. Calculation requires knowing which documents are relevant; it cannot be inferred from an open search across the whole web.
#
The harmonic mean of precision and recall. It summarises two dimensions but does not itself account for the severity of an error or a missed document.
#
A value calculated from a file to detect differences or check a copy. It establishes neither authorship nor content quality.
#
An ordered record of operations and decisions supporting examination of events. Check coverage, timestamps and log access.
#
The minimum insertions, deletions and substitutions needed to convert the reference to the recognised text, divided by the reference character count. Fix whitespace, accent and case rules before comparing OCR engines.
#
The technical ability to export and reuse data in another system. Check formats, attachments, metadata and reimport; downloading alone does not prove portability.
#
The ability to leave a tool or return to an earlier state with controlled effort. Verify it with a restoration trial rather than an export promise alone.
#
Operating software on infrastructure you manage. It requires maintenance and backups and does not guarantee all calls remain local.
#
Locations where data is stored or processed. The publisher’s country is insufficient: check regions, subprocessors and contractual documents.
#
Information that identifies a person directly or indirectly. Check necessity and processing terms before sharing it.
#
The period for which a service stores inputs, outputs or logs. It varies by product, plan and settings.
#
A model whose trained parameters are distributed under a license. It does not always mean the data or code is open.
#
Terms governing use, redistribution or modification of a model, tool or content. Check the exact license before commercial use.
#No term matches this search. Show all terms