Clarifying Document Context Usage in Clay Generations

Casper S. · 2025-09-10T08:10:35.646Z

Hey y'all! I'm looking for documentation on how the Document Context is working exactly. I feel like sometimes generated text properly abides by strict rules from the documentation, other times it completely disregards. How does Clay use these documents as contexts, exactly? We've been updating our system prompts to include the main rules from the context because it keeps forgetting or is inconsistent, but this makes the prompts incredibly long and both scalability and maintenance are hurting. Any ideas?

Clarifying Document Context Usage in Clay Generations | Clay

7 comments

· Sorted by Oldest

Natalie H.
·
·
Hi Caspe, thanks for reaching out! Since Document Context is still in beta, we don't currently have any documentation for it yet. Do you mind sending the link (url) to the table where the context was forgotten/inconsistent, so we can take a look? Thank you!

Loading attachments
Casper S.
·
·
https://app.clay.com/workspaces/492981/workbooks/wb_0t0l3no98igVZ3dAJNV/tables/t_0t0timwesAozQrF6WJN/views/gv_0t0timwFxDXAfc27Zwe
Casper S.
·
·
Can't tell you exactly which cells, seeing that there are very very many, though noticed in our random checks that there was a wide variety in ways of answering even though our brand guide is quite tight.
Casper S.
·
·
What we mostly saw:
Writing style (words, sentence builds, etc)
Reference to details from our research document in context
Sometimes hallucinated pains/gains from persona even though they are clearly listed in the context
Generally wondering how they work. When I address the documents through for instance a vector database (RAG setup) in custom n8n agent builds the output is much more predictable and follows writing rules and guidelines much more closely.
Casper S.
·
·
Understandable. It would be great having some form of understanding on the technical side how documents get weighed in context or if they get tokenized or whatnot. Currently it's too unpredictable, hence our ridiculously long system prompts. This black box limits rapid scaling Clay AI in niche applications a lot.
Casper S.
·
·
Without of course spreading detailed technology details. It'd be very helpful to know at least how the documents are considered by the AI on the other side. I assume it's basically an obfuscated API call, so knowing how those documents get sent as context is very relevant to many use cases. Especially when drilling to the core of prompting and when the need for high quality output is high.
Casper S.
·
·
It is, which is why i was wondering why the output is less consistent. I'll do some tests.

Natalie H.
·
·
Hi Caspe, thanks for reaching out! Since Document Context is still in beta, we don't currently have any documentation for it yet. Do you mind sending the link (url) to the table where the context was forgotten/inconsistent, so we can take a look? Thank you!

Loading attachments
Casper S.
·
·
https://app.clay.com/workspaces/492981/workbooks/wb_0t0l3no98igVZ3dAJNV/tables/t_0t0timwesAozQrF6WJN/views/gv_0t0timwFxDXAfc27Zwe
Casper S.
·
·
Can't tell you exactly which cells, seeing that there are very very many, though noticed in our random checks that there was a wide variety in ways of answering even though our brand guide is quite tight.
Casper S.
·
·
What we mostly saw:
Writing style (words, sentence builds, etc)
Reference to details from our research document in context
Sometimes hallucinated pains/gains from persona even though they are clearly listed in the context
Generally wondering how they work. When I address the documents through for instance a vector database (RAG setup) in custom n8n agent builds the output is much more predictable and follows writing rules and guidelines much more closely.
Casper S.
·
·
Understandable. It would be great having some form of understanding on the technical side how documents get weighed in context or if they get tokenized or whatnot. Currently it's too unpredictable, hence our ridiculously long system prompts. This black box limits rapid scaling Clay AI in niche applications a lot.
Casper S.
·
·
Without of course spreading detailed technology details. It'd be very helpful to know at least how the documents are considered by the AI on the other side. I assume it's basically an obfuscated API call, so knowing how those documents get sent as context is very relevant to many use cases. Especially when drilling to the core of prompting and when the need for high quality output is high.
Casper S.
·
·
It is, which is why i was wondering why the output is less consistent. I'll do some tests.