Use a prompt to look at a specific document in the vector store

Prompting
1

I have many files uploaded into the vector store. I have added one that contains a list of HSE activities and associated risks. sample:

1. Site Preparation and Setup
Activities:
Site clearance and demolition
Setting up temporary structures (e.g., scaffolding, barriers)
Ground marking and surveying
Associated Risks:
Demolition Hazards:
Structural collapse or falling debris
Exposure to asbestos or other hazardous materials
Scaffolding Risks:
Falls from height
Scaffold collapse or instability
Improper assembly leading to accidents
Site Clearance:
Musculoskeletal injuries from manual handling
Vehicle and machinery accidents
Surveying:
Trips, slips, and falls on uneven or cluttered ground
Exposure to extreme weather conditions
2. Excavation and Foundations
Activities:
Digging trenches and foundations
Pouring concrete
Installing formwork and reinforcements
Associated Risks:
Excavation Hazards:
Cave-ins and trench collapses
Contact with underground utilities (gas, electricity, water)
Exposure to hazardous atmospheres (e.g., confined spaces)
Concrete Work:
Burns from wet concrete (caustic hazards)
Respiratory issues from concrete dust
Heavy lifting injuries
Formwork Installation:
Falls from height
Structural failure of formwork
Manual handling injuries

In the System instructions I use:

Risk Assessment Prompt
Use the attached document 'UK Building Trade Activities and Associated HSE Risks.txt' to build a list of activities and associated risks with each aspect of the project.

Can the assistant ‘see’ the ‘UK Building Trade Activities and Associated HSE Risks.txt’ file and hence act on this.

So if the prompt contains 'I am building an extension ’ will it pull out the relevant activities and risks ?

Any help appreciated.

2

Uploaded documents placed into the vector store have text extracted, and are broken up into chunks of information, default 800 AI tokens in length.

The only way that an AI in assistants can interact with uploaded knowledge is by writing a search query, which is sent to retrieve those chunks that rank highest, to then be able to answer user questions that would hopefully be informed by some of that text.

There is no method for the AI to observe all the documents or even a single document at once (unless of short length and matching the search).

Thus, if you were to provide that overview, perhaps as part of your instructions that says “file_search has this kind of information in documents, which are your own proprietary knowledge…”, it might emit higher-quality searches, but the skill that gives would be to answer about specific questions in subject areas given.

Analysis of the entire corpus of regulations or guidelines would take building a retrieval platform of higher understanding - like a separate AI as tool that could act on your own code’s version of document search, to then parse through all relevant documentation (or ALL at great expense), looking for sections to extract which match the criteria.