All work

A Slack assistant that answers from your own documents

A Slack bot that drafts outreach in a founder's voice or answers internal questions, using the company's own documents as its source.

Client
Agency founder's sales team
My role
Built, tested and tuned
Period
Feb to Mar 2026
Status
Built and tested on realistic scenarios

Mock-up: made-up question and documents

51
steps in the assistant workflow
54
steps in the document-loading workflow
4
active AI agents, backed by 5 AI models
0.45
minimum relevance score for search results

The problem

The first version mixed in information from unrelated documents. I traced it to the search returning weak matches. I fixed it with a minimum relevance score, so weak matches are thrown away, and wrote the fix up so it could be repeated.

The main workflow

The assistant workflow (n8n)

SuccessErrorSuccessErrortruefalseSuccessErrorSuccessErrorSuccessErrortruefalse012SuccessErrorSuccessErrorSuccessErrorSuccessErrortruefalsetruefalseSuccessErrortruefalseFetch TopicsFetch TagsBuild Filter PayloadSlack TriggerNeeds Reranking?Extract ScoresSet QueryBuild Rerank PayloadRerank (Cohere)Parse PineconeResultsBuild AI Payload(Reranked)Build AI Payload(Pinecone)Embed QuerySearch PineconeSlack NormalizerOutput Parser:IntentSet Query MetadataQuery TaggerChat ModelOutput ParserSearch Pinecone(Fallback)Has Matches?LLM: Founder(gpt-5.2)LLM: Intent Router(gpt-5-mini)Route by ModeVA Info AgentChat ModelMemoryLLM: Query Tagger(gpt-4.1-mini)Send Slack ReplyEnd (No-Op)Founder Agent (WithContext)Chat ModelMemoryLLM: FounderFallback (gpt-5.2)Intent RouterChat ModelOutput ParserLLM: VA Agent(gpt-4.1-mini)Output Parser:Query TaggerSet ContextResponse NormalizerMemory: FounderMemory: VAEval TriggerFounder Agent (NoContext)(Deactivated)VA RAG Agent(Deactivated)Skip Rated RowsWrite Response toSheet(Deactivated)Eval NormalizerHas Thread?Fetch Thread HistoryFormat ThreadContextNo Thread(Passthrough)Dedup CheckIs Query Empty?Send Empty ReplyValidate ResponseLog Delivery Failure
Drawn from the real workflow export in the style of the n8n editor. 51 working steps (n8n calls them nodes). Drag to move around, or use the zoom buttons.

What I built

  • A 51-step workflow that first works out what the person is asking for. It runs three more AI agents (one that labels the question, one that drafts outreach and one that answers internal questions), remembers the Slack thread, ignores duplicates and checks each reply before sending it.
  • A separate 54-step workflow that loads new documents: it watches a Google Drive folder, cleans PDF, Word and subtitle files, cuts them into pieces, labels each piece with AI, saves them to Pinecone and posts progress in Slack.
  • Search through Pinecone, with Cohere re-sorting the best matches to the top and a minimum relevance score of 0.45, so weak matches never reach the AI.
  • A Google Sheets test set with scripted scenarios and simulated users.
  • A 400-line guide that documents the fix for made-up answers and how to apply it elsewhere.

Tools used

  • n8n
  • Slack
  • Pinecone
  • Cohere rerank
  • OpenAI
  • Google Drive
  • Google Sheets