Turn Slack Chats into Searchable Knowledge with n8n and Local AI

Slack channels are where most team knowledge actually lives. But when someone asks "did we decide about X?" six months later, the answer is buried in a thread nobody can find. n8n plus local AI can capture, summarise, and index those discussions into a searchable knowledge base automatically.

What to capture

Not every Slack message is worth indexing. Focus on:

• **Decisions** — threads where a clear conclusion was reached
• **Technical Q&A** — how-to questions with good answers
• **Architecture discussions** — design rationale and trade-off notes
• **Process updates** — changes to workflows, policies, or tooling
• **Project milestones** — completion announcements and retrospectives

For the knowledge base side of this setup, see Build a Team Knowledge Base with n8n and AnythingLLM.

Architecture

```

Slack Events API → n8n Webhook → Filter → Classify (local LLM) → Summarise → Index

↓ ↓

Low signal / noise Send to knowledge base

(skip or archive) + store metadata

```

Step-by-step build

Step 1: Set up the Slack integration

Create a Slack app with these event subscriptions:

• `message.channels` — listen to public channel messages
• `message.groups` — listen to private channel messages (if desired)
• `message.im` — direct messages (use with caution)

Set the webhook URL to your n8n webhook endpoint. Use Slack's signing secret to verify requests.

Step 2: Filter for signal

Most Slack traffic is noise for knowledge purposes. Add a filter node in n8n that passes through only messages that:

• Have more than 3 replies (indicates discussion, not broadcast)
• Contain a question mark (likely Q&A)
• Have reactions like :white_check_mark: or :done: (indicate resolution)
• Come from designated "decision" or "knowledge" channels

Step 3: Classify with a local LLM

Pass qualifying thread content through an AI Agent node:

> Classify this Slack thread. Is it: decision, technical_qa, architecture, process_update, or noise? If noise, set skip=true. If useful, provide a one-sentence summary. Return JSON.

Step 4: Summarise and index

For threads classified as useful, generate a structured summary:

| Field | Source |

|-------|--------|

| Title | LLM-generated from thread topic |

| Summary | LLM-generated 2-3 sentence overview |

| Decision | The conclusion or answer, if applicable |

| Participants | Slack user IDs from the thread |

| Channel | Origin channel name |

| Date | Thread date |

| Link | Permanent Slack link to thread |

| Tags | LLM-generated keywords |

Send the structured record to AnythingLLM or your preferred knowledge base via its API.

Step 5: Periodic digest

Create a weekly n8n workflow that compiles captured knowledge into a digest:

1. Query the knowledge base for entries added this week

2. Group by channel or topic

3. Generate a summary using local LLM

4. Post the digest back to a designated Slack channel

Privacy and permissions

Be transparent with the team about what is being indexed. Provide an opt-out mechanism for sensitive channels or message types. Never index private conversations without explicit consent.

For the privacy fundamentals, read How to Secure a Self-Hosted AI Server.

Maintaining the knowledge base

Indexed Slack knowledge needs periodic maintenance:

• **Monthly review** — archive outdated decisions or superseded information
• **Cross-reference** — link related threads that discuss the same topic
• **Stale detection** — flag entries older than 12 months for review

Conclusion

Slack contains a wealth of team knowledge that is invisible to formal documentation tools. n8n and local AI make that knowledge accessible by capturing, classifying, and indexing discussions as they happen. The result is a knowledge base that grows organically without requiring anyone to "document things properly."

FAQ

Will this capture every Slack message?

Only messages that pass the signal filter. You control what gets indexed via the classification prompt and channel allowlist.

How much does this cost to run?

Nearly zero if using local models. The main cost is the compute time for classification and summarisation.

Can this work with Discord?

Yes. Discord has similar webhook and bot API patterns. The same n8n workflow design applies with a different trigger node.