Combining tables with text in Prompt node

[sohanasarah]

I have some pdf files and also some table data. I am trying to use Promptnode with a DPR retriever. How can I add the table data with the text files? I'm not using text table retriever because it's not giving me good results, and I want the results to be in Long Form.

[ZanSara]

Hey @sohanasarah, table data is not easy to work with for text-based models. If the tables are large enough, table retriever is your best option right now. Otherwise you may try to leave the tables in the documents, or process them as text, and see if the results are any better. Unfortunately there aren't any other simple solutions to this problem...

[sohanasarah]

Thank you very much @ZanSara. I'm using prompt_node for generative LFQA. Still not getting good results. My retriever can retrieve the correct documents, but the result is incorrect when I use the prompt node in the pipeline. I am using the Google flan t5 xl model with prompt node. Any suggestion on how to improve the results? Also, how can I finetune the prompt texts in the custom prompt template? Couldn't find any guidelines anywhere.

[ZanSara]

Regarding the accuracy of answers, unfortunately that mostly depends on model size: if your retriever works well and still promptnode fails to understand what it should reply to your questions, you either fine-tune the prompt (but that works only on a narrow set of questions usually) or you use a more powerful model (which normally just makes the problem disappear).

Regarding the prompt text: PromptTemplare objects have a prompt_text field that you can set to any string you want. If you have trouble figuring out how to do so, share the part of your code where you create the PromptNode and I'll try to help.

[sohanasarah]

  ```      
        retriever = EmbeddingRetriever(document_store=document_store_2,
                                       embedding_model='chatbot\downloaded_models\all_datasets_v3_mpnet-base',
                                       model_format="sentence_transformers",
                                       use_gpu = True)

        lfqa_prompt = PromptTemplate(name="lfqa",
                             prompt_text="""Synthesize a comprehensive answer from the context for the given question. 
                             Your answer should be in your own words and be no longer than 1000 words. 
                             \n\n Context: {join(documents)} \n\n Question: {query} \n\n Answer:""",
                             output_parser=AnswerParser()) 
        prompt_node = PromptNode(model_name_or_path="google/flan-t5-xl",
                                 default_prompt_template=lfqa_prompt,
                                 use_gpu=True)

this is the code I am using. 

[sohanasarah]

I am using some pdfs that contain texts, equations, and tables. I tried both haystack pdf converters and separately converting them to JSON objects and then add as document object. I checked a few queries with the document search pipeline.