Giving your AI assistant sources: what research shows
Does an AI assistant answer better when given company documents? Retrieval-augmented generation research offers a reasoned answer.
The principle
Patrick Lewis and co-authors (2020, Advances in Neural Information Processing Systems) combine a language model with document retrieval: on factual question tasks, answers are more accurate.
Breaking down reasoning
Jason Wei and co-authors (2022, Advances in Neural Information Processing Systems) show asking the model to spell out steps improves some reasoning tasks, mainly for large models.
The nuance
These are academic benchmarks; quality depends heavily on the documents. An outdated document yields an outdated answer. (our executive and employee training programmes)
The limit of common practice
Loading every departmental file into the assistant, unsorted and undated.
A practical check in Bern
Restrict the assistant to ten validated, dated documents and test twenty real questions with known answers. The accuracy rate is the baseline before expanding. To go further: explore the Building AI Assistants and Agents Without Coding training in Bern, or browse our executive and employee training programmes in Switzerland.
In pictures: Building AI Assistants and Agents Without Coding in Bern
- building ai assistants and agents without coding
- Bern
- research
- training bern