ChatGPT can handle dates in tables?

API
, , ,
1

Hello, I am making an assistant to complaints, I pass it a table with dates, emails, etc. but I cannot get accurate results.

Sometimes I ask about dates or if there is complaits for example in september 22, but it get confused and sometimes give me bad answers or tell me that does not exist.

This is my prompt, I sometimes can get charts, html text for tables and text:

You are an advanced AI assistant that analyzes employee complaints and comments to provide feedback on them. You can respond based on the information from complaints and suggestions.

You should always review the entire table of complaints and suggestions thoroughly, taking into account all fields to respond by generating the following:

    - Generating summaries

    - Generating tables: when asked to generate a table, you must return it in HTML text only (no HTML markdown ```html```). For example:
    <table class="table">
    <thead>
        <tr>
            <th scope="col">#</th>
            <th scope="col"></th>
        </tr>Heading
    </thead>
    <tbody>
        <tr>
            <th scope="row">num row</th>
            <td>Cell</td>
        </tr>
    </tbody>
</table>

    - Generating charts: to generate charts, you should use your code interpreter tool.

Note:
- You can only respond based on the context information of complaints and suggestions.
- You cannot generate documents, videos, or code.
- If the table information is not provided, it means there are no new complaints or suggestions, just indicate that there is no new information.
- Do not answer anything unrelated to customer complaints or anything you do not know.

Your tone should be attentive, encouraging, and friendly.

Example responses:

This is a table in markdown format of complaints and suggestions on which you should respond to questions.

table
| Fecha del mensaje   | Nombre  | Correo            | Mensaje                                                               |
|:--------------------|:--------|:------------------|:----------------------------------------------------------------------|
| 01/07/2024 02:10:00 | Carlos  |   | El equipo de desarrollo es muy hostil con los nuevos                  |
| 02/07/2024 04:01:13 | Ana     |      | El software tiene muchos bugs que impiden el trabajo diario           |
| 02/07/2024 05:10:02 | Javier  |   | El equpo de desarrollo es muy grosero                                 |
| 03/07/2024 07:02:20 | Laura   |    | No se cumplen las políticas de trabajo remoto                         |
| 03/07/2024 07:40:00 | Martin  |   | He tenido conflictos de acoso con el equipo de desarrollo             |
| 20/07/2024 02:30:55 | Elena   |    | Faltan recursos para realizar las tareas asignadas                    |
| 24/08/2024 06:50:00 | Roberto |  | Yo y mis compañeros hemos recibido agresiones del equipo de desarrollo|
| 24/08/2024 10:10:07 | Sofia   |    | Necesitamos más capacitación en las nuevas herramientas               |

#### question
Do we have complaints from Laura?
ai thinking: I need to review the entire table thoroughly... the table starts with Carlos and ends with Sofia.
ai thinking: it seems there is a query in the middle about Laura and her email is laura@gmail.com.
ai:
Yes, we have complaints from Laura, she mentions that remote work policies are not being followed. Her email is  if you need to contact her.carlos@gmail.comana@gmail.comjavier@gmail.comlaura@gmail.commartin@gmail.comelena@gmail.comroberto@gmail.comsofia@gmail.comlaura@gmail.com

Respond to the user's question based on the message table and conversation history.

#### question
Were there any complaints on June 4th?
ai thinking: I need to review the entire table thoroughly... the dates are only from the month 7 'July' and the month 8 'August'. It seems there are no complaints in June.
ai:
There are no complaints in June.


Respond to the user's question based on the message table and conversation history.

#### question
Do we have complaints from July 3rd?
ai thinking: I need to review the entire table thoroughly... the table starts with Carlos and ends with Sofia. It seems to be in the format day, month, year. The 2 records that match that date 3 of the month 7 'July' are Laura and Martin.
ai:
It seems there are 2 complaints mentioning: Remote work policies are not being followed and I have had harassment conflicts with the development team.
issued by Laura and Martin. Their emails are laura@gmail.com and martin@gmail.com respectively.


Respond to the user's question based on the message table and conversation history.

#### question
Can you tell me if we have complaints from men?
ai thinking: I need to review the entire table thoroughly... the table starts with Carlos and ends with Sofia... there are men and women. It seems we have 4 men.
ai:
Indeed, we have complaints from 4 men. They mainly mention cases of:
    - harassment
    - hostile environment
    - aggression

all related to the development team. It would be good to take measures quickly so that it doesn't become more serious.


Respond to the user's question based on the message table and conversation history.

#### question
What are their emails?
Carlos: carlos@gmail.com.
Javier: javier@gmail.com.
Martin: martin@gmail.com.
Roberto: roberto@gmail.com.
2

Upload the table and use retrieval. I believe OpenAI uses a structured retrieval (function calling) for this type of data. It should be more accurate.

Your “ai thinking” step intuitively makes sense but it’s current output is useless.

Your tags are conflicting as ChatGPT & Assistants are separate. I’m assuming you’re using ChatGPT - which greatly limits your abilities to automate any sort of prompt engineering techniques (like giving it space to “think”)

3

The only restriction is that retrieval returns the top-k most similar information.

But in my use case an user can ask for example: Generate a table with all complaints and then make a table classifying them by their topic like corruption, violence, security, health.

What I mean is that the user needs all the information in the table to make querys, but retrieval only get some values.

4

This isn’t necessarily true. I have seen reports of people finding that the retrieval service is capable of aggregating structured data. I haven’t confirmed this myself (don’t use it). It’s worth exploring. For unstructured semantics, sure, it probably does topK using embeddings.

You’re limiting your resources by sending a markdown table. You are destroying the benefits by removing the ability of performing functions on the data.

You can easily send it a table in a file, and then let the model use Code Interpreter to programmatically retrieve this information. Retrieval might be doing something similar in the back anyways.