Evidence & Safety

Clinical reasoning support, with the veterinarian in control

Terav is an AI clinical reasoning assistant for veterinary professionals. It helps organize case information, interpret laboratory findings, explore differential diagnoses, and identify questions for further investigation.

Terav brings together original veterinary literature, structured laboratory rules, and AI-assisted reasoning. Its literature retrieval gives the AI access to relevant passages from original articles, with links back to the sources for clinical review. The veterinarian remains responsible for assessing the patient and deciding which investigations and treatments are appropriate.

This page explains how these mechanisms work and where their limits lie. Source retrieval and some additional checks depend on the workflow and configuration. They are not applied identically to every response.

Original articles as the evidence base

Terav’s article knowledge base contains the original text of source publications. The text is preserved without rewriting or summarization and divided into smaller passages, called chunks, so relevant material can be found for a clinical question. Those passages remain associated with their source articles.

Terav uses retrieval-augmented generation, or RAG, to bring this material into the conversation. When relevant passages are retrieved, the AI receives the original article text as context alongside the available case information. Source links let you return to the publications and examine their findings in context.

This separates two parts of the process: the evidence supplied to the AI consists of original source passages; the explanation that applies those passages to the case is generated by the AI. Retrieving article text does not mean the underlying model was trained on those articles.

The knowledge-management workflow includes preparation, structural validation, review and approval, and indexing of approved material. Source records and document versions support maintenance and traceability of the collection.

Some workflows also support a separate web-retrieval route. When enabled, this route filters external material using domain-based authority rules and can require material from more than one domain before passing it to the AI. These checks govern that additional retrieval route; they do not describe how original article text is preserved in the RAG collection.

Retrieval is selective. It is not a systematic literature review and does not guarantee complete coverage, the latest guidance, or applicability to a particular patient. If retrieval is unavailable or does not return useful material, some workflows can still generate a response without retrieved evidence.

From a source passage to a clinical interpretation

Source links accompanying retrieved article material point back to the original publications. They allow you to inspect the evidence, read the surrounding discussion, and assess whether the study population and findings are relevant to your patient.

Preserving the original text provides a direct foundation for an answer. The remaining limitation concerns how the AI selects, interprets, and applies that evidence: it can overlook a qualification, draw a conclusion broader than the passage supports, or apply a finding to a patient whose circumstances differ from those studied. A retrieved passage also represents only part of an article’s full context.

Terav has additional citation checks in specific workflows. The laboratory interpretation pipeline checks citation presence and format and compares source names against a known-source list. Certain response-processing paths also check citation identifiers against the supplied source set. These checks complement source retrieval; they do not independently establish that every generated clinical claim follows from the cited passage.

Responses may also contain model-generated bibliographic references. These require verification and should be distinguished from source links attached to retrieved original articles. Not every statement receives an individual evidence check.

For a decision that depends on a particular claim, use the source link to review the original article, including its methods, population, limitations, and relevance to your patient. The source provides the evidence; applying it to the individual case remains a clinical judgment.

Laboratory checks before AI interpretation

In the structured laboratory workflow, a deterministic rule engine evaluates numeric results before the AI generates its interpretation. It normalizes test names, applies supported unit conversions, compares values with available reference intervals, and checks for predefined patterns and critical values.

Where a usable reference interval is extracted from the uploaded report, the rule engine gives it priority over a catalogue reference. Catalogue coverage and matching depend on the test and patient information available.

These checks provide a structured starting point. They are not a comprehensive clinical assessment. Unsupported tests, missing references, unusual units, incomplete context, and extraction errors can limit the result. A value within a reference interval does not establish that the patient is healthy.

The AI receives the rule-engine output as context. This does not guarantee that every sentence it generates is consistent with that output.

Reviewing extracted results

Reading a PDF or photograph can introduce errors in values, test names, and units. Terav checks extracted results for issues such as uncertain recognition, unfamiliar test names, unit mismatches, and implausible numeric values.

When these checks flag results for review, the structured laboratory workflow pauses before interpretation. You can correct flagged entries or confirm that they match the original report. Outstanding flags must be resolved before that review can be confirmed. The confirmed data are then used for interpretation.

Results that do not trigger a flag can proceed automatically. This is not a guarantee that extraction was error-free. Compare the extracted information with the original report, particularly when a finding will change clinical management.

Making uncertainty visible

Terav is designed to keep a differential diagnosis open. Its clinical reasoning format asks for evidence for and against hypotheses, missing information, and diagnostic steps linked to the hypotheses they would help assess.

The support labels used in this format mean:

• Supported: several independent findings align with the hypothesis.
• Possible: the hypothesis is compatible with the case, but information is limited or a competing explanation remains.
• Unlikely: the hypothesis fits poorly but has not been excluded.

These labels are qualitative assessments generated by the AI. They are not calibrated probabilities. Internal confidence scores also use heuristic signals and should not be interpreted as the probability that a diagnosis or recommendation is correct.

Missing information and confident wording both require attention: an AI response can be wrong even when it sounds decisive.

Keeping the clinician involved

Clinical review responses include a clinical check that invites you to compare the assessment with your own findings. Challenge prompts let you ask Terav to argue against its leading hypothesis, examine a finding it may be overemphasizing, or identify what would change its assessment.

These features support active scrutiny. They do not provide independent specialist review or require formal clinician approval of every generated answer.

Terav may discuss diagnostic approaches and treatment options. Any suggestion requires your assessment of the patient, contraindications, relevant guidance, and clinical priorities. Terav is not intended to issue a final diagnosis or make prescribing decisions autonomously.

When information or services are unavailable

Different problems produce different responses. Flagged laboratory extraction can require your review before interpretation continues. If the AI interpretation step fails, the laboratory pipeline can return a limited rule-based result. If source retrieval fails, a response may still be generated without that additional context. If an optional rewriting step fails its checks, the system can retain the earlier response.

These fallback behaviours preserve parts of the workflow; they do not certify the remaining output as safe or complete. Do not treat the absence of a warning, reference, or review request as evidence that every relevant risk has been checked.

Terav is not an emergency monitoring or triage service. A response or an urgency label must not delay direct assessment and necessary patient care.

Traceability and feedback

Terav stores case messages and generation metadata that can help investigate how a response was produced. Depending on the workflow, metadata can include the model and provider, processing route, timing, source links, and whether retrieval or additional processing was used.

Operational AI-call records capture identifiers, timing, success or failure, and a prompt hash rather than a dedicated full prompt-and-response transcript. These records are separate from stored case conversations. Knowledge-management actions also have administrative audit records.

This supports troubleshooting and review. It is not a complete, immutable clinical audit trail or a guarantee that an earlier answer can be reproduced exactly.

You can rate a response and add a comment to report an issue. Feedback supports product review; it does not automatically correct the case or validate the answer.

What Terav does not establish

Terav does not examine the patient, observe facts that have not been supplied, or replace veterinary clinical judgment. Its safeguards do not guarantee detection of every extraction error, unsupported claim, contraindication, or urgent condition.

This page makes no claim of clinically validated diagnostic accuracy, improved patient outcomes, or regulatory certification. Source checks and software tests should not be read as evidence of those outcomes.

Terav is intended for veterinary professionals, not as a self-diagnosis or treatment service for pet owners. Use its output as material to review alongside the patient, the original records, and appropriate clinical evidence.