Skip to main content

Uploads

Supported types are PDF, Word (.docx), Markdown, text, CSV and JSON. Convert other formats (for example .doc, .pptx, .html) to one of these first.
The PDF is probably scanned (an image of text). Truvec reads the text layer and doesn’t run OCR. Convert it to a searchable PDF and upload it again.
The file has no extractable text. Check it opens and contains selectable text.
Lower the overlap or raise the chunk size. A common setting is an overlap of 10 to 20% of the size.
Select the programming language of the files in the Language field.
The JSON strategy only works on JSON files. Use Recursive for other file types.
Pre-chunked uploads must be a .json array, a .jsonl file with one chunk per line, or a .csv file with a text column. The error names the item or line to fix: a missing or empty text, a repeated id, or invalid JSON. See supported formats.
Other documents of the same file type were chunked with different settings. Choose Use existing settings unless you mean to mix them. See Chunking consistency.

Golden datasets

Upload documents before creating a golden dataset.
The AI model couldn’t produce usable questions from your documents, or it was temporarily unavailable. Try again, with fewer questions if it keeps failing. Documents with very little text (scans, tables only) give the model little to work with.
One of its expected chunks was deleted with its document. The question can no longer be answered correctly. Create a new dataset, or remove the question before committing.
Only committed datasets can be evaluated. Open the dataset and click Commit Dataset.

Evaluations and comparisons

Open it in Evaluation History to read the error. The most common causes are:
  • The embedding provider was temporarily unavailable. Run it again.
  • The golden dataset has no entries.
  • Your plan’s embedding volume is used up for this period. The error says how much is left and when it resets.
The first run with a model embeds all your chunks, which can take a few minutes for large projects. Later runs are much faster. If a run is still pending or running after an hour, email support@truvec.dev: we investigate, then re-run it for you or refund it.
The model’s or the reranker’s price wasn’t known when the run happened. Costs are computed at run time, so run the evaluation again to get them. See Model pricing.
The evaluation was run with an earlier version of Truvec. Run it again to inspect each question.
Documents were added or removed after the run. The results describe the old documents. Re-run the evaluation, and review the golden dataset if you added documents.
The reranker’s provider can’t be reached from Truvec right now, so runs that use it are refused before they start. Pick another reranker, run without one, or try again later. The Reranker list only shows the rerankers that are available.
The documents were deleted or chunked again after the golden set was built, so some of its expected chunks are gone and the scores are wrong. This always happens after a chunking change. Generate a new golden set on the current chunks. See Comparing chunking strategies.
A reranker only reorders the first-stage candidates. If the right chunk wasn’t among them, it can’t bring it back. Open the question in the Retrieval Inspector: if the expected chunk is listed under Expected but not retrieved, raise First-stage candidates, try hybrid search, or a better embedding model.

Projects and account

If deletion fails with “Delete project evaluations first”, delete the project’s evaluations and comparisons, then delete the project again.
Sessions last 2 hours. Sign in again.