Knowledge Corpus Curator

Review the files used by an AI knowledge source for duplication, redundancy, staleness, overlap, and potentially conflicting guidance. A cleaner, better-governed knowledge corpus can improve grounded response quality, reduce retrieval noise and response latency, and make content easier to maintain.

The skill analyzes uploaded copies without changing the source files. It combines all uploaded batches into one corpus and returns an Excel review backlog, an HTML summary, JSON results, and a batch manifest.

The Excel backlog and HTML report include a shared UTC creation timestamp in their filenames, such as knowledge-corpus-curation-backlog-2026-07-22-142530Z.xlsx, so repeated runs do not overwrite earlier reports.

The Excel backlog consistently contains four tabs: Review Backlog, Summary, Document Inventory, and Curation Settings.

Before you start

• Download the SharePoint library or folders you want reviewed.
• Include only files you are authorized to process.
• Optionally add the corresponding SharePoint site or library as agent knowledge to help validate high-risk findings.

Agent knowledge is optional. The uploaded files remain the authoritative analysis corpus. This workflow does not require SharePoint connector actions or a SharePoint MCP server.

What SharePoint knowledge enables

Adding the corresponding SharePoint site, library, or document location as an agent knowledge source gives the skill an additional validation layer alongside the uploaded files. It can:

• Check whether current SharePoint passages support or clarify high-risk duplicate and conflict findings.
• Identify audience, region, effective-date, or scope differences that may explain an apparent conflict.
• Find related or potentially authoritative documents that were not included in the uploaded corpus.
• Flag relevant SharePoint documents as Not included in uploaded corpus for a future review batch.

SharePoint knowledge does not replace the uploaded source files. Retrieved knowledge passages are partial chunks, so they cannot prove whole-document duplication, a complete-document conflict, or exhaustive library coverage. They also do not reliably provide original owners, URLs, approval status, or modified dates; supply the optional metadata manifest when those fields matter.

Prepare the files

Package the export as one or more independent ZIP files while preserving the SharePoint folder structure when possible.

• Each uploaded ZIP can be up to 50 MB.
• The combined size of all files uploaded in one chat session can be up to 200 MB.
• Target 45 MB or less per ZIP to leave practical headroom.
• Use logical batches such as business area, policy family, document type, or top-level SharePoint folder.
• Keep each document in only one batch.
• Do not use split archives such as .zip.001; every ZIP must open independently.

If the intended corpus exceeds 200 MB, reduce the scope or analyze separate corpora in separate chat sessions. Results from separate sessions are not automatically combined, so cross-session duplicate, similarity, and conflict coverage is incomplete.

How to use it

1. Add the skill to your Copilot Studio agent.
2. Attach one or more ZIP files. Multiple ZIPs can be attached to the same message, and additional batches can be sent in later messages in the same conversation. Each ZIP can be up to 50 MB, provided the combined size of all files attached in the chat session remains under 200 MB.
3. If the agent asks whether more files are coming, confirm that the upload is complete or attach another batch. No specific confirmation phrase is required.
4. The agent stages all confirmed uploads together, analyzes every uploaded file, validates the highest-risk findings, and returns the report files.

Final-batch confirmation is the only required intake question. The skill does not ask whether the upload is a whole library or subset, and it does not filter drafts, archives, or historical versions. Every file in the confirmed upload is included. Staleness uses the default 365-day threshold unless you already requested another value.

What it finds

• Exact and normalized-text duplicates
• Near-duplicates that may drift over time
• Related or overlapping content
• Potentially conflicting guidance with paired evidence
• Stale-content candidates
• Unsupported files and extraction gaps
• Relevant knowledge-source files missing from the uploaded corpus

Potential conflicts and duplicate findings include the related filenames, locations, excerpts, and PDF page numbers where available. Recommendations are review actions only; the skill never deletes, moves, renames, publishes, archives, or overwrites source content.

Confidence values are deterministic numeric detection scores where available. Blank confidence means the analyzer did not produce a score; it never means a finding was human validated.

The backlog is always ordered by priority: Critical, High, Medium, then Low, so the most urgent review items appear first in every output format.

Columns remain consistent across runs. The Excel backlog retains primary and related paths for traceability, while the HTML report omits those path columns for readability. The report title includes its UTC creation date and time, and the source ZIP filename or filenames appear directly below the title.

Good to know

• Upload all batches in one conversation so comparisons work across batch boundaries.
• Files from an earlier conversation should not be assumed to remain available.
• SharePoint downloads may replace original modified dates. Supply the optional metadata JSON described by the skill when accurate freshness dates, owners, URLs, or approval status matter.
• Large corpora can exceed the configured pairwise comparison threshold. Exact duplicate detection remains corpus-wide, but semantic comparisons may be limited and will be disclosed in the report.
• If the runtime does not include the Excel dependency, the skill still returns HTML and JSON, and the JSON file contains the complete backlog.