Knowledge Base and Internal Q&A

The document library is the foundation. Quality here determines quality everywhere else. Spend 2 - 4 weeks on this before touching n8n. Identify the categories of reference material that generate the most internal questions: engagement management policies, methodology documentation, conflict of interest procedures, billing guidelines, template library, and precedent files for common situations. For each category, identify the single authoritative current version of each document. Archive or delete outdated versions - having multiple versions of a document in the indexed library is worse than having none, because the system might return the outdated version. Format documents consistently. Long PDF documents with complex formatting are harder for the indexing system to process accurately than clean text or Markdown files. For critical policy documents, create a clean text version even if the formatted PDF exists for external use. Build a document registry: a simple spreadsheet listing every document in the library, its category, the date it was last reviewed, and who owns it.

Install the n8n vector database node for your chosen platform. The indexing workflow reads each document from your library, breaks it into chunks (typically 500 - 1000 word segments with some overlap), generates embeddings for each chunk, and stores the embeddings with metadata (document name, section, date, URL/file path) in the vector database. Run the indexing workflow once initially to index the entire library. Set up a trigger to re-index any document that's been updated - this keeps the knowledge base current as documents are revised. After initial indexing, test the retrieval by querying for your 20 most common questions. For each question, review which document sections were retrieved. Are they the right sections? If not, adjust the chunking strategy or the indexing approach before moving to the answer generation step.

The Q&A bot workflow triggers when a message is sent to a designated Slack channel or tagged with a specific keyword. Set up the Slack app or Teams bot following your platform's documentation - both have extensive guides for creating bots that respond to messages. When a question arrives, the workflow: 1. Sends an immediate acknowledgment (typing indicator or "Looking that up...") 2. Generates an embedding for the question using the same embedding model used during indexing 3. Queries the vector database for the most relevant document chunks (top 3 - 5 results) 4. Passes the question and retrieved chunks to the AI for answer synthesis 5. Returns the answer with explicit citations to the source documents For questions where the retrieval doesn't find relevant content (similarity score below threshold), return: "I don't have a clear answer to that in the current knowledge base. I've flagged this as a gap for the document team - in the meantime, [suggested escalation path]."

You are an internal knowledge assistant for a professional services firm. Your job is to answer questions from team members using only the document excerpts provided - you must not use information from outside these excerpts.

Question: {{user_question}}

Relevant document excerpts retrieved:
{{retrieved_chunks}}

(Each excerpt includes: document name, section, and content)

Instructions:
1. Answer the question directly and specifically using the provided excerpts
2. If the excerpts clearly and fully answer the question: provide the answer, then cite the source(s) in this format: "Source: [Document Name], [Section], [Date if available]"
3. If the excerpts partially answer the question: provide what you can, note the limitation, and cite what you used
4. If the excerpts don't adequately answer the question: say so clearly - "The current knowledge base doesn't have a clear answer to this question. The closest relevant content is [brief description] in [document]. I've flagged this as a knowledge gap."
5. Never fabricate information. If you're uncertain, say so.
6. Keep answers concise - aim for under 200 words unless the question requires more detail

Format your response as:
**Answer:** [Direct answer]
**Source(s):** [Citations]
**Note:** [Any important caveats or limitations, if applicable]

Every question the system can't answer well is a signal that the knowledge base needs a new document or a clearer policy. Build a gap tracking workflow: when the system returns a "not found" or low-confidence answer, log the original question, the date, and the requester to a gap tracking sheet. Assign document owners to review the gap log weekly and determine which gaps represent genuine policy gaps (need a new document written), unclear existing policies (existing document needs clarification), or out-of-scope questions (appropriately not in the knowledge base). Set a monthly review cadence: document owners review their sections, update content that's changed, and archive anything that's outdated. When a document is updated, the re-indexing workflow triggers automatically.