Context Engineering Data
Let us look at following interaction.
I visit my usual restaurant to celebrate my birthday. Upon entering, here is the conversation.
- Host: Good afternoon sir, Happy Birthday.
- Me: Thank you.
- Host: The usual place sir?
- Me: Yes please.
The host takes me to the table with a good view of TV. Favorite is game is on today.
- Server: Good afternoon sir, Happy Birthday.
- Me: Thank you.
- Server: What would you like to have today?
- Me: The usual.
- Server: May I remind you that Shravan, the holy month is going on.
- Me: Oh Yes. Thanks for that. Please replace Chicken with Paneer.
- Server: Anything else sir?
- Me: No. Thanks.
- Server: Most welcome sir and Happy birthday to you again.
After a few minutes, I get Paneer Butter Masala extra hot, with Roti and Rice. Mango milkshake with soy milk, as I am lactose intolerance.
Get Kulkarnisatishp’s stories in your inbox
Join Medium for free to get updates from this writer.
Few minutes into my eating, restaurant manager notices me and comes over the table to greet me with happy birthday. We chat a little about the food and service and he asks for ratings and recommendations when done. Then he leaves me to continue with food and the game.
Let us break the interaction down in terms of context.
1. Persona and data points.
2. How deep the data should be.
3. Who can see what?
Persona and data points
First question is what to capture. I came across the concept of persona while working with one of my clients so I will continue using it. I have the same data as I am me. But every persona looks me from different angles.
Host: Knows me, knows that today is my birthday and what my favorite place to sit is.
Server: Knows me, knows that today is my birthday and what my favorite food is.
Chef: Knows me, knows food customizations I want (extra spicy and soy milk).
Manager: Knows most about me including birthday, favorite place to sit, usual order and customization. He may use the same during conversation with me. However, he is more inclined towards my view on restaurant food and service at high level.
How deep the data should be
This is interesting question for me. How much data the sever need to define “The usual” based on my past interaction? I may be ordering the same dish there or may different dishes over years. There could be a pattern that I order same dish every birthday, there could be a pattern that I order certain dishes when I am alone and certain dishes when I am with family. (Note: my family is vegetarian) or when I am hosting a party. Pinpointing an exact “Usual” is what is called as golden nugget in context engineering.
Who can see what
Every persona have their own POV which mandates certain data to be available to them. Should we maintain separate data for persona or should we allow filtering of data based on the persona? This becomes very complicated when we start to compress the data to save space in terms of catering to a context window.
Do you think this covers most critical aspects of context engineering data? Anything I have missed, do let me know.
Note: I have explicitly kept the implementation details out of scope such as data storage and compression techniques, Access control mechanism.



![dbt Incremental Models: How to Pick the Right Strategy — [dbt Series #9]](https://miro.medium.com/v2/resize:fit:1358/format:webp/1*RGVFr99XbaklPI-E2DEeYw.png)



