Voyage AI publishes an API for creating vectors and a model for reordering found fragments. These steps can be measured separately.
The second stage sees only the list of candidates
A help search ranks the correct page seventh. The pages above it use similar words but describe another version. Voyage AI rerankers accept a question and a list of passages and return scores. You can add them after vector retrieval or another first stage. They see only that list, so they cannot help when the right page was left out.
How to Measure Change
On the training documentation, prepare questions with previously specified sources. Write down the first results before the second call and after. Look not only at the average position, but also at the queries where the correct fragment has gone down. For example, the old section could describe the same command in more detail. Then it is useful to transfer the features of the version or remove the old edition before ranking.
Check the number of candidates. The longer edition gives the second model more choice, but adds text to the query and can increase latency.
Connection to a job search
Refine permission to transfer source fragments through the API. Save the model ID and rules for preparing the login next to the results of the check. When changing the terms of payment, recalculate costs separately for indexing, searching and re-ranking. If the processing is not suitable for some documents, exclude them before an external call and check that the user still receives a clear result.
Suitable tasks
- Select the order of fragments after the primary search
- Check the cost and delay of the second stage
Strengths
- Document Embeddings and Ranking as Separate Challenges
- You can work with the issuance of an existing index
Limitations
- Reranking cannot recover candidates omitted by the first retrieval stage
- Verify that you have permission before sending documents to an external API
Selection and conditions
- Free access
- Not confirmed
- Trial period
- Not confirmed
- No sign-up required
- Not confirmed
- Russian interface
- Not confirmed
- Local installation
- Not confirmed
- API
- Yes
“Not confirmed” means there is not enough information. The feature may still be available; check the developer’s terms.
- Input formats
- Document, Text
- Result
- Data
Terms
Embeddings and reranking follow separate billing rules. Check the service terms and limits before connecting. Reranking cannot recover candidates omitted by the first retrieval stage.From task to first result
Practical example
- Task
- Select the order of fragments after the primary search
- What to prepare
- A few search questions, candidates found and known correct sources.
Task brief or prompt
Pass the question and found candidates for re-ranking. Compare the first results with the initial results.
Expected result
The order of documents before and after processing and the list of requests with a changed result.
How to check the result
- Note the questions for which the candidate was not needed.
- Measure the delay and limit the number of transmitted fragments.
An editorial teaching example of a task brief, not the result of hands-on testing of the service. Check that the required mode is available, that you have rights to the source material and that the intended use of the result is permitted.
Check details
Official sources
This description is based on the developer’s material. Features, prices and availability may change; check the primary source before use.