How to add table annotations serialization.

[kupelabs]

Question

...

Hi everyone,

I have gone through the docling documentation and found this article table exportation to get the different tables based on which I wrote a code for an LLM to generate a caption (though getting some context for example text data regarding the page from which the table is found could really be a plus). I will like to do something similar to the picture annotation serialization provided here image annotation serialization but for the generated table captions (annotations). I noticed that the TableItems objects have a captions field. Any help is appreciated thanks.

[dosubot]

You can serialize table annotations (including generated captions) in Docling by creating a custom serializer, similar to how image annotation serialization works. The TableItems objects' captions field is designed to store these annotations, and you can extend the serialization logic to include them in your output.

While there isn't a direct example for table caption serialization in the docs, the process is meant to parallel the image annotation approach. You can define a custom serializer that extracts the captions from TableItems and includes them in your serialized output. For reference, see the custom serializer example for images and adapt it for tables.

If you want to include additional context (like page text), you'll need to extend your serializer to access the relevant document model fields and output them alongside the table and its caption. The Docling system is set up for this kind of extensibility, so you can follow the same pattern as with images, but target the TableItems and their captions field instead. More on customizing serialization can be found here.

If you need a code template or more specific guidance, let me know what language or framework you're using and I can help you sketch it out!

Written by @dosu, approved by dolfim-ibm

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[dolfim-ibm]

Adding to the bot answer. We are generalizing the concept of annotations from pictures to more objects (including tables). So we should soon have a way to populate the annotations using llms and to add them in the exported docs.