i have created conversational agent and uploaded and updated all documents and folders and agent in proper folders.
i have created for company policy so if i ask question related to company policy then agent can give reply.
i have both word and pdf documents and tried with strategy “semantic” and “structured(preview)” for context.
issue is it is not able to search in proper document, and every time asks for escalation approach even though i have specified all in system prompt and question is valid as answer is there in document.
what should i do to work my conversational agent work as expected?
Hey Kunal,
Please share the screenshots of the issue, along with the prompt, the query, and context documents you are using.
Also, need more clarification on this- i have created conversational agent and uploaded and updated all documents and folders and agent in proper folders.
Let’s figure this out
Looking at your agent’s response, I believe it’s working fine partially (as in it is suggesting to escalate the query to HR team as mentioned in the prompt).
First option - Could you try removing all other context-related documents apart from the one which has information related to “time of process” and reframe the query a bit, like “time taken to complete XYZ process end-to-end”?
Second option- Check for any other query the answer to which is there in the same context document?
Third option- Just have the leave policy document as part of the context and query that
if i only put only required document then what is the use of conversational agent after specifying all the instructions ?instead i directly use chatgpt.
i will try to keep required document only and try (just to validate working or not)
as per my question replicon word only contains in one document out of 3
my question - what is the time off process in replicon
documents in storage bucket - check below ss
Checked your prompt and here are some suggestion and observations:
Prompt is long and instructions given are ambiguous
Lack of instructions on how the agent should interpret the documents or the content in the document
Goal (first line in the prompt) does not cover entire responsibility of agent
Suggestions for improvement:
Give clear and complete goal description
Guidelines (conversation and behaviour) can be in single section to avoid duplicity and ambiguity
Knowledge source - can mention exact or identical files names with short description of document usage
Escalation steps and tool usage steps mixed and not concise
As per my observation it looks like prompt is written manually, instead per my experience I would suggest to give instructions to autopilot and it’ll write better prompt. Prompt refinement and improvement also you can do it from autopilot itself to ensure consistency.
I this help or improvise your solution or help in resolution please mark as solution so it can help others as well.
could you please help me with sample promt? because i have done that also as you have suggested but not worked u can check attached promt here HR policy assistant system promt2.txt (2.3 KB)
consider i have 10 policy documents for my company and agent need to search in those documents and identify proper document based on query and need to give reply only from that document.
as it is company related documents so agent should not give answer from google or using only gpt it should only from docs and if not found then it should go for escalation approach
