Knowledge Seeding

How to bulk-ingest your accumulated professional knowledge into Coppermind so it starts smart instead of learning from scratch.

Why Seed Knowledge?#

Over years of practice, a fractional CMO accumulates frameworks, playbooks, templates, vendor evaluations, industry benchmarks, and methodology documents. Without seeding, Coppermind only knows what you tell it one conversation at a time.

The tune tool lets you point to a folder of your professional materials and bulk-ingest everything at once. An LLM classifies each file as general CMO knowledge versus client-specific, then routes it to the right client mind automatically.

The Two-Step Process#

Knowledge seeding is always a dry-run-first workflow. This prevents accidental ingestion of the wrong files.

Step 1: Preview (Dry Run)#

Point to your knowledge folder and let Coppermind classify everything without ingesting:

"Scan my knowledge folder at ~/Documents/CMO-Library for ingestion"

Coppermind calls tune with dry_run=true (the default). You get back a classification summary:

Found 47 files. Classification:
  31 general CMO knowledge (frameworks, templates, benchmarks)
  12 client-specific (8 matched to existing minds, 4 unmatched)
   4 skipped (duplicates, binary files)

Review the classification above, then re-run with dry_run=False to ingest.

Each file shows its classification reason and knowledge type, so you can spot misclassifications before committing.

Step 2: Ingest#

Once you are satisfied with the classification:

"Go ahead and ingest my CMO knowledge folder"

Coppermind re-runs with dry_run=false. Each file goes through the standard extraction pipeline (chunk, extract, dedup, reconsolidate) and lands in the correct client mind.

Where Files Go#

The LLM classifies each file into one of four categories:

General knowledge goes to your personal mind (created automatically if it does not exist). This is CMO-level training -- reusable knowledge that helps across all client engagements.

Supported File Types#

The tool processes the same file types as folder ingestion:

• .txt -- plain text
• .md -- Markdown
• .docx -- Word documents
• .pdf -- PDF files

Binary files, images, and unsupported formats are classified as "skip" automatically.

Parameters#

The tune tool accepts four parameters:

Overriding Auto-Classification#

If you want everything to go to one client mind regardless of classification:

"Ingest everything from ~/Documents/Acme-Archive into the Acme Corp client mind"

This uses the target_mind parameter to bypass classification routing.

If you only want general knowledge (no client routing):

"Seed my knowledge from ~/Documents/CMO-Library but skip client-specific routing"

This uses auto_route_clients=false -- client-specific files go to your personal mind instead of being routed.

Deduplication#

Re-running the tool on the same folder is safe. Each file is hashed before ingestion, and files that have already been ingested are skipped. The ingestion receipt shows how many files were deduped:

Ingestion complete. 12 files ingested, 35 already ingested (deduped), 0 skipped.

This means you can add new files to your knowledge folder over time and re-run without duplicating existing content.

When to Use It#

Typical First-Time Flow#

1. "Scan ~/Documents/CMO-Library for knowledge seeding"
   → Review classification summary

2. "The Acme-specific files look correct, but file vendor-list.md is general, not client-specific. Go ahead and ingest."
   → Coppermind ingests with the noted correction

3. "I also have an archive for an old client, Bluebell. Create a client mind for Bluebell and then re-run."
   → Creates the client mind, re-runs, catches the previously-unmatched files

Edge Cases#

• No personal mind exists: Created automatically on first ingestion. No action needed.
• All files are client-specific: Nothing goes to the personal mind. That is fine -- not an error.
• Unmatched client files: Files that clearly belong to a client but do not match any existing client mind are flagged in the summary. Create the client mind and re-run to pick them up.
• Large folders (100+ files): Files are classified in batches to stay within LLM context limits. Processing time scales linearly.
• Extraction failures: Files that cannot be read (corrupted, password-protected) are logged in the error list and skipped.

How It Relates to Other Features#

• Remembering one thing ("Remember: ...") is for individual notes. Knowledge seeding is for bulk loading many at once.
• Ingesting a transcript is for meetings. Knowledge seeding is for documents and files.
• Your personal mind is set up automatically the first time you seed knowledge into it.
• Search -- after seeding, your personal mind is richer, and searches across everything you seeded work immediately.
• Morning briefing (/cmo-brief) can reference seeded knowledge (e.g., "You have a template for quarterly reviews" or "You've worked with this vendor before").