solvi documents

Ask typed questions of a document and get answers you can check. A ModernBERT extractor and the solvi decision engine run entirely in your browser.

Extractor: not loaded
790 MB, downloaded once and cached by your browser.
Python (Pyodide) + solvi: loading…
Decision logic runs here, in WebAssembly.

Document


        

      


      

        

          

            
Fields (defined by description)

            
          

          

            
namedescriptionfound

            
          

          

            A new field is extracted as soon as you leave the description box.

          

        


        

          
Answers

          
Answers appear here after the fields are extracted.

        

      


      

        

          
Rules (Python, runs in your browser)

          

            
            
          

        

        
Every field above is a fact with the same name (a string, "" when absent).
          @cat.fn computes a fact, @cat.check(hard=True, then={...}) is a check that overrides the
          answer when false, @cat.rule("q") answers question q. Helpers: money,
          number, parse_date, days, hours, found,
          need, mentions, iban_ok. INPUTS adds inputs; today
          is always one.

        
        

      


      

        

          
How the answer was reached

          

            
            
            
          

        

        

        
Run a use case to see the strategist's flow, every computed fact and the hash-chained trace.

      


      

        
An honest note on the model

        
solvi-ai/extract-base
          is a general starting point, trained on contracts, receipts and Wikipedia-style questions. Fields close to that
          work well out of the box; fields far from it can be missed or cited at the wrong place, and you will see it in the
          highlight. For production, label 25–100 of your documents and fine-tune (LongSpanExtractor.fit,
          see the model card). The rules, checks, abstentions and the trace do not depend on the model being right: they are
          what lets you see when it is not.

        
Add a use case

        
A use case is one file in usecases/: a title, fields as [name, description], sample
          documents and the rule code. Copy usecases/receipt.js, change it, list it in
          usecases/index.js and open a pull request on
          GitHub.

        
What leaves your browser

        
Downloads only: the model and tokenizer from Hugging Face, Python (Pyodide) and onnxruntime-web from jsDelivr,
          and the solvi package from PyPI. Documents, fields and rules stay in this tab (edits are kept in your
          browser's local storage so they survive a reload).