An automated IT knowledge-base builder. On a schedule, it scans OneDrive for IT documentation, classifies every document with OpenAI, and builds cleaned, deduplicated, AI-enriched Notion databases, one per IT category.

Built as a Pipedream workflow with seven custom Node.js steps. All step code is in steps/.

Category routing

Whatever label the AI or a source assigns, each record lands in one of ten Notion databases. Records whose category has no configured database are reported per item instead of failing the run.

Features

• Scheduled scanning: recursively walks a OneDrive folder tree through the Microsoft Graph API on Monday, Wednesday and Friday at 11:00 UTC, with an exclude list and a file cap.
• AI classification: GPT-4o returns strict JSON for each document (type, key topics, technical relevance, use cases, category and a confidence score).
• Deduplication: merges OneDrive, Excel and OneNote records and removes near-duplicates by similarity, keeping the better-extracted copy. Dedup state can persist across runs in a Pipedream Data Store.
• Category routing: 27 aliases (azure, identity, virtual desktop, …) map onto ten databases: Network Infrastructure, VDI, Applications, Hardware, Security, MDM, Cloud Services, Backup Recovery, User Management and Monitoring.
• Normalization: one standard schema with ISO-8601 dates, tags split on five delimiters, and canonical status and priority values. Unrecognized fields are kept under metadata.
• AI enhancement: a second OpenAI pass writes titles, summaries, tags, priority and optional custom sections such as action_items, with a fallback when the model returns bad JSON.
• Notion sync: matches enhanced records back to their pages, rewrites content blocks, and can re-parent pages and archive ones that no longer have a matching record.

Stack

• Pipedream workflow (Node.js, ES modules)
• Microsoft Graph API for OneDrive
• OpenAI (GPT-4o for scanning, gpt-3.5-turbo by default for enhancement)
• Notion API
• GitHub Actions syntax check for every step

Quickstart

1. Open the share link: pipedream.com/new?h=tch_BXf12z
2. Connect your Microsoft OneDrive, OpenAI and Notion accounts when prompted.
3. Point the scanner at a OneDrive folder and pick the file types to include.
4. Create or map the Notion databases for the categories you want, and set the schedule. This deployment runs 0 11 * * 1,3,5 (UTC).

How it works

1 · OneDrive knowledge scanner

steps/01_comprehensive_onedrive_processor.js

Walks the folder tree (/items/{id}/children), filters by MIME type (Word, PDF, plain text, Markdown, OneNote) and extracts content. Plain text and Markdown are read in full; binary formats are captured as structured metadata descriptors. Each document goes to GPT-4o for classification, near-duplicates are dropped by Jaccard word-set similarity, and the step emits one of three output shapes (json, summary, knowledge_base) with per-run metrics.

2 · Clean and deduplicate

steps/02_clean_and_deduplicate_data.js

Merges records from up to three sources (OneDrive scan, Excel exports, OneNote content), normalizes text, and removes duplicates by percentage similarity in configurable batches.

3 · Create specialized IT databases

steps/03_create_specialized_it_databases.js

Creates Notion databases at the workspace level, which avoids failures from archived parent pages. Properties are defined as JSON and support title, select, multi-select, number, date, people, files, checkbox, URL, email, phone, relation and rollup types.

4 · Populate by IT category

steps/04_populate_specialized_databases.js

Routes each record to its database through the category map, builds full page properties and body blocks, and reports created pages and per-category counts.

5 · Normalize records

steps/05_normalize_database_records.js

Resolves field names case-insensitively against candidate lists (title/name/subject/…), normalizes second and millisecond timestamps to ISO-8601, and maps status and priority values to canonical vocabularies ("done" → completed, "blocker" → critical).

6 · AI enhancement

steps/06_enhance_individual_records.js

Produces an enhanced title, detailed summary, categories, suggested tags, priority and content type per record, plus any caller-defined custom sections.

7 · Update Notion pages

steps/07_update_notion_pages_with_enhanced_content.js

Matches records to pages by page ID, title or original ID, updates properties, and appends or replaces formatted content blocks.

Design notes

• No secrets in code. Authentication goes through Pipedream's managed app connections (microsoft_onedrive, openai, notion). Nothing in this repo holds credentials.
• Fail-soft. Every step collects per-item errors instead of throwing, so one bad document never stops a run.
• Configurable. Similarity thresholds, batch sizes, file caps, output formats and custom AI sections are exposed as step props.
• Known limitation: PDF, Word and OneNote content is analyzed from metadata descriptors, not full binary text extraction. A document-parsing service such as Azure Document Intelligence is the natural next step.

Project layout

steps/          source of each workflow step, exported from Pipedream
docs/images/    README diagrams
.github/        CI (syntax check for every step)

License

Copyright (c) 2026 Isaiah Anson. All rights reserved. You may use the released software for personal, non-commercial use; copying, modifying or redistributing it requires written permission. See LICENSE.