AI-Powered Document Finder in Shareflex QMS

The Shareflex QMS cockpit gives access to all approved quality documents. The list view includes powerful filters and columns to quickly search and organize content.

 

While experienced IT users appreciate this interface, less tech-savvy users often struggle to find the documents they need.

In this post, we show how to add a ChatGPT-like interface to your Shareflex QMS that helps users find the most relevant valid documents — using smart AI-powered matching instead of old-school full-text search. Users can ask natural questions, even with typos, and still get accurate results.

Here are some example search queries users might enter:

  • How do we report a security incident?
  • What’s the procedure for locking the building?
  • Wat moet ik doen bij brand?
  • email hack what to do?

Even with spelling mistakes or vague wording, the system understands the intent and shows matching documents.

Let’s give a demo of how it looks and feels (later in this post, we will explain how it was created).

The user asked the question “Give me all IT procedures”

The Open document link will open the procedure in Shareflex QMS.

Some more questions (question is in right bottom corner field)

Now that we used typo errors in the input, we still get the results we were looking for: inappropiate behevior

The question is in Polish, while the content is in English – yet we still get a match.

How the database was made

We used Make.com scenarios and various bundles to create the data processing workflows.

With below displayed scenario a special Pinecone vector database has been created to store the data needed to answer the end user questions on the QMS documents

List Files from SharePoint – Retrieves all files in the designated SharePoint folder where approved QMS documents are stored. Acts as the starting point.

Get SharePoint Metadata – Fetches detailed metadata per file, including file name, path, SharePoint ID, listItemId, and more — used for linking back later.

Convert File to Plain Text – Uses CloudConvert to convert PDF/Word files into plain text, which is required for summarization and embedding.

Clean Text (Module 6) – Removes formatting noise and line breaks. Ensures the text is optimized for OpenAI input.

Summarize with GPT-4o –Uses a Chat Completion call to GPT-4o to summarize the policy/procedure. Focuses only on core meaning, rules, and guidance (max 4 lines).

Generate Embedding – Sends the cleaned text to OpenAI’s text-embedding-ada-002 API to generate a semantic vector.

Upsert in Pinecone – Stores the embedding vector along with metadata into the Pinecone ShareflexQMS namespace:

title (from file), summary (from GPT), file URL / listItemId, doc type, validUntil, etc.,

 

How the chat user interface was created

Retool was used to create the user chatbot. 

User enters a question in a Retool text field.

User clicks the “Ask” button.

Retool sends a POST request to a Make.com webhook with the question and optional result limit.

Make.com processes the request (embedding, vector search, GPT filter).

Make returns matching documents as text.

Retool displays the result in a text box.

The Make.com scenario that answers the Retool user queries is shown below.

Allow users to ask natural-language questions (in Retool), and return the best-matching valid QMS documents from SharePoint—powered by OpenAI embeddings and Pinecone vector search.

Parts Involved

Retool Webhook Input – A Retool app sends the user’s question via webhook to Make.com. Input: question (user text), limit (number of matches to return).

Clean Input Text – Replaces line breaks with spaces to make clean embedding input.

Create Embedding – Sends cleaned question to OpenAI (text-embedding-ada-002) to generate an embedding vector.

Pinecone Vector Search – Searches the ShareflexQMS namespace using the generated vector. Returns the top N matches (including metadata and vector scores).

Relevance Filter via ChatGPT – For each match, asks GPT-4o: “Does this document answer the user question?” Prompt includes question + document metadata (summary/title/type). GPT replies: YES, xx% or NO, xx%.

Text Aggregation – Filters out all NO matches. For YES matches, builds a formatted list: #3 – Clean Desk Policy – Valid until: 08/31/2026 Don’t leave confidential info unattended.

Return to Retool – Sends the aggregated results back as a simple text block for Retool to show in the chat UI.