Bad responses on a longer conversation

[guanjiahui]

@makeshn @trebedea

The config file: instructions:

• type: general content: | Below is a conversation between a bot and a user where the user asks the bot questions about a patients health and the bot answers the questions based on documents the bot has access to. The user can ask as many questions to the bot. Any question that does not pertain to the data the bot has access to should not be answered.

sample_conversation: | user "Hello" express greeting bot express greeting "Hello" user "What does the patient present with?" ask question bot answer question with sources "Patient presents with Paracentesis" user "Summarize past patient problems" ask question bot answer question with sources "The patient has a history of Paracentesis, with cirrhosis and ascites"

models:

• type: main engine: openai model: gpt-3.5-turbo-instruct
• type: embeddings engine: openai model: text-embedding-ada-002

[drazvan]

@guanjiahui: what's the exact issue?

[guanjiahui]

Hi @drazvan, I shared this issue during DLI with @makeshn. Basically, we got bad answers after two questions. The answer includes contents that are related the the previous two questions. See highlights in the pic