I am a relatively new GPT user( 2-3 months)

I am trying to leverage the power of ChatGPT to do the following:
Receive and parse a CSV BOM(could also be XLSX instead) - this could be small to medium qty of BOM items, ranging from 10 items to 900, and retrieve the Product URLs & Specifications/Datasheet URLs if available.

The CSV BOM will have these columns:
Quantity
Manufacturer
Description
Product URL (to be enriched via GPT)
Specifications URL (Optional - to be enriched via GPT - could be the same as the Product URL)
Datasheet URL (Optional - to be enriched via GPT)

Required process flow:

1. review file name for hygiene

2)aggregate the BOM items to Unique Product IDs with QTY sum

3. group BOM items by manufacturer

4. for each manufacturer, process the BOM by unique Product IDs
   a) confirm part number in BOM via searching to find authoritative product page (complicated by the fact that some different product brands are now same company/family/distributor due to mergers/acquisitions)
   b) enrich the sheet with the Product URL
   c) scrape the Product URL for specifications and/or links to Datasheet
   d) enrich the sheet with Specifications URL( Could be the product page)
   e) enrich the sheet with Datasheet URL (optional - may not exist)
   f) cycle through BOM items not yet with a resolved defensible URL that loads a real page for user inputs of a product page URL or snip with link and Product ID & Manufacturer

Output in XLSX or CSV to be the original uploaded CSV

I have tried this:
via simple prompts after uploading a CSV
via creating a custom GPT with structures SOT
-File naming Convention.docx (enforces hygiene for traceability)
-Output Template.XLSX
-GPT Instruction Protocol.docx

Both routes end in sadness
-Way to much user input required. i may as well have just searched and found the URLS myself)
-GPT blatant disregard for hardcoded instruction in the GPT Instruction Protocol.docx
a) using foreign language symbology in the output against instructions
b) regularly hallucinating untruths
c) Indicating completion but nowhere near
d) arbitrary undisclosed decisions to omit BOM items
e) upon completion, creating whole other separate work flow for me to incorporate(install Python, run a script) to create the output required that the whole exercise was to complete

The sophistry is killing me

Any suggestions/Guidance?

maybe this is wrong place, i have since joined the discord and and waiting to be able to post there

Welcome to the community!

I just wanted to jump in and say that I think this is a reasonable place to ask, since it relates to ChatGPT workflows, file handling, prompting and avoiding hallucinated results.

That said, from what you describe and if I’m understanding you correctly, this may need more than ChatGPT alone, likely some external tooling or automation around the spreadsheet/search/validation part.

There are users here who are much more experienced with those kinds of workflows, so hopefully someone can point you toward the right setup.

Thanks,
Adjacent to that, the tool is for myself, but may be shared with other users, so trying to design with that in mind. Not everybody is going to have python skills( i dont yet, but i will), let alone the environment to run without jumping through hoops to get IT to allow new software installed

aisystemsdev:

If I were you:

1. Create a project
2. Describe exactly that in a project.
3. Save to project things you like from chat.
4. eventually , you may want to formalize concepts into “system_architecture.md”files
5. Depending on how you like to build projects you might just need to ask the AI to simply take that as a build spec and build a python program which accepts a csv input and outputs as requested.
6. Or if you’re like me, you may want to document specs first and plan out the project before starting to write code. It really depends on if you need a simple program or if you’re designing a larger project.

—

I think the core issue is that this is not really a “prompt harder” problem.

You’re asking ChatGPT to do several different jobs at once:

• clean and validate an uploaded BOM
• normalize duplicate part numbers and quantities
• search the web
• decide which result is authoritative
• avoid mergers/distributor confusion
• scrape or inspect product pages
• find datasheets
• verify that URLs actually load
• preserve every original row
• output a clean CSV/XLSX without omissions

That is a lot of moving parts, and the failures you’re seeing are exactly where GPTs tend to struggle: silent omissions, overconfidence, unverifiable URL choices, and “I finished” when only part of the workflow is actually done.

I would probably break this into a project/workflow instead of expecting one prompt or one Custom GPT to do the whole thing end-to-end.

For example:

• Step 1: validate filename and column structure
• Step 2: aggregate unique Manufacturer + Part/Product ID pairs
• Step 3: create a “research queue” of unresolved parts
• Step 4: research only one manufacturer batch at a time
• Step 5: require evidence for every URL decision
• Step 6: mark each item as Resolved / Needs Review / Not Found
• Step 7: only then merge results back into the original BOM

The key is not asking it to “complete the BOM,” but asking it to produce an auditable enrichment table where every decision has a status and source. I would also explicitly forbid it from dropping rows and require a final reconciliation count:

• original row count
• unique item count
• resolved count
• unresolved count
• final output row count

For this kind of task, I would not trust a single Custom GPT instruction document to enforce everything. I’d use the GPT more like a supervised research assistant with checkpoints. If the workflow becomes repetitive enough, then yes, it may eventually justify a small tool or script, but the design should come first: input format, status fields, validation rules, reconciliation rules, and what counts as a defensible URL.

In short: I’d stop trying to make the GPT magically enrich the whole file in one pass. Make it produce a traceable research queue, work in batches, require evidence, and force reconciliation before output.

ChatGPT alone is useless for this type of operation.

My advise is to use Codex. There are some Codex gurus who may be of help.

Looks like you are involved in an MRP project. Been there - many moons ago.

If so, get ready for a wild ride as the complexities are enormous:

BOM (Bill of Materials) is the hierarchical “recipe” listing all the raw materials, subassemblies, and components required to build a finished product.
MRP (Material Requirements Planning) is the software system that uses that recipe, combined with inventory data and sales forecasts, to calculate what to buy or manufacture, how much to buy, and when to order it.

How They Work Together

The relationship between BOM and MRP is linear and interdependent in manufacturing. The MRP relies entirely on the BOM to break down a finished product into its individual parts.

1. Demand & Schedule: The process starts with the Master Production Schedule (MPS), which dictates what finished products need to be built and by when.
2. The Recipe (BOM): The MRP engine “reads” the BOM for that finished product to understand the exact quantities of parts needed.
3. The Inventory Check: The MRP checks your current inventory status to see what is already on hand or already ordered.
4. The Plan: Taking lead times into account, the MRP generates a schedule and purchasing orders to fill the gap between what you have and what you need to meet the deadline

If that’s his problem, then absolutely this would require either an agent with persistent memory, but preferably a python program to handle the logic and ensure accuracy. Personally; I wouldn’t trust an AI because over time there will be drift and it’s more double checking than you’d want especially for batch workflows.

I’ve designed multiple in house solutions for both myself and friends who needed consistent data driven results. If you laid out your requirements plus sample data to work with to refine the process I could build this out for them. Or anyone could given the use of AI agents and direction to design the program.