Why OpenAIAnswerGenerator gives answers out of context?

[nachoperezzv]

Hi!

I'm using the (retriever + answer_generator) QA pipeline with the Davinci-003 (openAI) model as generator. My problem is that whenever I give my retriever a question that doesn't fit with any document It still gives me the most similar document. The given context is fed to the answer_generator as well as the question. Now, my question is: Is there any way to force the generator to give an answer that is only related to the given context ?

Example:

question:
Hello, what's your name?

document_text from retriever:
Arya is introduced as the daughter of Eddard Stark and Catelyn Tully. She is shown as a willful, stubborn girl with character and who challenges the feminine roles of her time. In the HBO television adaptation, Game of Thrones, she is played by actress Maisie Williams.
`

Now, the answer generated by the davinci-003 model is something like:
My name is ...

By I'd like that the answer fits with the given context. For example, something acceptable would be:
`I can not give you an answer with the given context'

I don't know if there's a way of adjusting the answer with some parameters like the temperature, or the frequence_penalty

[anakin87]

I'm just giving you an opinion and maybe it's not the best idea...

• What are the scores returned by the retriever for unrelated documents?
  If these scores are low (as I hope), you can hand-craft a mechanism that doesn't call OpenAIAnswerGenerator in these cases, it just returns something like
  I cannot give you an answer with the given context

• Another option would involve the PromptNode, which already has a question-answering-check template.
  It would try to figure out if the answer is contained in the context...

[foochris]

@nachoperezzv I was able to find a solution by passing a custom prompt template to the OpenAIAnswerGenerator:

my_template = PromptTemplate(
    name="qa-more-reliable-answers",
    prompt_text="Please answer the question according to the above context."
    "If the question cannot be answered from the context, reply with 'I cannot find the required information in the context.'"
    "\n===\nContext: {examples_context}\n===\n{examples}\n\n"
    "===\nContext: {context}\n===\n{query}",
)

# intialize GPT model
generator = OpenAIAnswerGenerator(api_key=api_key,
                                  model="text-davinci-003",
                                  max_tokens=50,
                                  presence_penalty=0.1,
                                  frequency_penalty=0.1,
                                  top_k=3,
                                  temperature=0.9, prompt_template=my_template)

The solution is documented here:

https://docs.haystack.deepset.ai/docs/answer_generator#customizing-the-prompt

[anakin87]

@bilgeyucel that useful prompt is a bit buried inside the docs. It might be useful to add it to the prompthub. WDYT?

[bilgeyucel]

@anakin87 Yes. Let's put it in the hub!

[foochris]

@bilgeyucel Can you create a PR (or issue) so the suggestion from @anakin87 doesn't get lost / forgotten?

[bilgeyucel]

Hi @foochris, I created an issue: deepset-ai/prompthub#38 👍