Aviagen Technical Assistant
Aviagen's technical publications contain management guidance and detailed performance and nutrition tables. Answering a question accurately means finding the right product, edition, age and unit, then explaining the result without losing that context.
VRG's work on the technical assistant combines document processing, structured numerical queries and AI responses. The engineering also covers evaluation, source inspection and usage accounting, so the team can investigate incorrect answers and understand the work involved in producing them.
Give different questions the right evidence
A numerical lookup needs a precise row and column. A management question may need several passages from a handbook. The assistant separates those paths rather than treating every request as a broad document search.
Supported numerical questions use structured data and deterministic formatting. Other questions retrieve relevant passages before generating an explanation. Product and edition checks help keep related publications from being mixed together.
Preserve tables and units
Some answer problems begin during document ingestion. Flattening a table can separate a number from its heading, row label or unit. A model cannot reliably recover relationships that were lost during extraction.
The ingestion workflow preserves document metadata, separates prose from numerical tables and normalizes supported tables into structured schemas. Targeted re-extraction and checks against source pages help repair damaged data.
Shared metric definitions keep qualifications such as daily versus cumulative values, or amounts per bird, attached to the answer. Missing verified data remains a visible limitation rather than a reason to substitute a nearby number.
Control the work behind each answer
The application bounds retrieved context, uses direct response paths where the data supports them and limits particular repair stages. Model selection is configurable by task, including routing, answer generation and evaluation.
Usage accounting records model operations and token consumption. Evaluation can reuse saved answers when comparing judges, avoiding another full generation run for every experiment. These are implemented cost controls; we do not claim a measured overall savings percentage.
Investigate failures and test the correction
Feedback replays retain the preceding conversation, which matters when a follow-up relies on an earlier product or age. The review process separates extraction failures, incorrect retrieval, answer defects and faulty evaluation expectations.
Citations link answers to document titles, sections and source PDFs where available. Readers can inspect the evidence while the engineering team uses regression tests to check fixes and responder precedence.
This case study describes engineering reviewed in September 2026. Individual benchmark runs are diagnostic evidence, not a guarantee of production accuracy. Current model choices and runtime configuration can change as the system develops.
Planning a similar project?
Explore how we scope and build AI applications, integrations and the software around them.
