Build an AI Knowledge Base for WhatsApp Auto-Replies

Your AI assistant is only as smart as what you feed it. Upload your docs and watch it answer questions with YOUR expertise — not generic ChatGPT knowledge.

MoltFlow's RAG search uses YOUR documents — not generic AI training data. Your AI knows your product as well as your best support agent. This guide shows you how to upload documents, process them for RAG WhatsApp AI retrieval, and test the AI knowledge base results.

What You'll Need

Before starting, make sure you have:

• A Business plan subscription — AI knowledge base (Business plan) — stop answering the same questions 50 times a day. RAG search and AI auto-reply features are available on the Business plan. Upgrade here if you're on a different plan.
• Documents to upload — PDFs, TXT files, DOCX files, or Markdown. Think FAQs, product catalogs, pricing sheets, service descriptions, or company policies.
• OpenAI API key configured — The AI knowledge base uses OpenAI embeddings for semantic search. Add your key in Settings > AI Configuration if you haven't already.

Good candidates for upload: support documentation that answers "How do I...?" questions, product specs, pricing policies, shipping information, return policies.

How the Knowledge Base Works

MoltFlow uses Retrieval-Augmented Generation (RAG) to give AI auto-reply responses factual grounding:

1. Upload & Chunk — When you upload a document, MoltFlow extracts the text and splits it into semantic chunks (roughly 500 tokens each, about 2-3 paragraphs).
2. Embed — Each chunk is converted into a vector embedding using OpenAI's text-embedding-ada-002 model. These embeddings capture the meaning of the text.
3. Search — When a customer asks a question, MoltFlow's RAG WhatsApp AI searches for the most relevant chunks using vector similarity.
4. Generate — The top matching chunks are fed into GPT-4 as context, along with the customer's question. GPT generates an AI auto-reply grounded in YOUR content.

The result: answers that come from your documentation, not hallucinated information. If your FAQ says "Shipping takes 3-5 business days," the AI knowledge base will respond exactly that — not "usually 1-2 weeks" or some made-up timeline.

Step 1: Navigate to the Knowledge Base

From your MoltFlow dashboard:

1. Click AI in the left sidebar
2. Select Knowledge Base

You'll see the knowledge base management page, which displays:

• Uploaded documents with their processing status (processing, ready, error)
• Chunk count for each document (how many pieces the document was split into)
• Upload date and file size
• Actions — download, delete, re-process

If this is your first time here, the page will be empty. Let's fix that.

Step 2: Upload Your First Document

Click the "Upload Document" button in the top-right corner.

A file picker will appear. Select a document from your computer.

Supported formats:

• PDF (text-based PDFs only — scanned images won't work without OCR)
• TXT (plain text files)
• DOCX (Microsoft Word documents)
• MD (Markdown files)

Max file size: 10MB per file

Best content to upload:

• FAQs — Highest ROI. If customers frequently ask "What are your hours?" or "How do I reset my password?", upload an FAQ document.
• Product catalogs — Detailed product descriptions, features, specifications.
• Pricing sheets — Plans, pricing tiers, what's included in each.
• Service descriptions — What you offer, how it works, turnaround times.
• Company policies — Refund policies, terms of service, privacy policy.

Click Upload and the file will be sent to MoltFlow's processing queue.

Step 3: Wait for Processing

Processing happens in three stages:

1. Text extraction (5-10 seconds) — MoltFlow extracts plain text from your PDF/DOCX.
2. Chunking (5-10 seconds) — The text is split into semantic chunks of ~500 tokens each. The chunker tries to split at paragraph boundaries and headings to keep related content together.
3. Embedding (30-60 seconds) — Each chunk is sent to OpenAI's embedding API to create vector representations.

You'll see the status update in real-time:

• Processing (orange badge) — Extraction and chunking in progress
• Ready (green badge) — Processing complete, document is searchable
• Error (red badge) — Something went wrong (usually API key issue or unsupported file format)

For a typical 10-page FAQ document, processing takes about 60-90 seconds total.

How Many Chunks?

A 10-page FAQ might create 20-30 chunks. A 50-page product catalog might create 100-150 chunks. More chunks = more granular search, but also more storage.

You can see the exact chunk count next to each document after processing completes.

Step 4: Test AI Replies

Now that your knowledge base is populated, let's test whether the AI can retrieve and use the information.

Enable AI Auto-Reply on a Test Chat

1. Go to Dashboard > Sessions
2. Click on your connected session
3. Go to Chats tab
4. Find a test chat (or create one by sending yourself a message from another phone)
5. Click the chat to open details
6. Enable AI Auto-Reply toggle

Send a Test Question

From the other phone (or have a friend help), send a message that your uploaded document should answer.

For example, if you uploaded an FAQ that includes "Our business hours are 9 AM to 5 PM EST," send the question:

"What are your business hours?"

The AI should respond with something like:

"Our business hours are 9 AM to 5 PM EST, Monday through Friday."

If the AI gives a generic answer like "I don't have that information," the document may still be processing, or the question isn't covered in your knowledge base.

Verify the Source

In the MoltFlow dashboard, go to AI > Conversations to see the AI's response and which knowledge base chunks were used. You'll see the retrieved chunks highlighted.

This helps you verify that the AI is pulling from your content, not hallucinating.

Step 5: Manage Your Documents

Update a Document

If you need to update information (e.g., new pricing, changed hours), upload the new version:

1. Click Upload Document again
2. Select the updated file
3. Give it the same filename (MoltFlow will detect the duplicate and ask if you want to replace)
4. Confirm replacement

The old chunks are deleted and the new chunks are created. This takes another 60-90 seconds.

Delete Outdated Documents

If a document is no longer relevant:

1. Click the trash icon next to the document
2. Confirm deletion

All chunks and embeddings are permanently deleted. The AI will no longer retrieve information from this document.

View Chunk Details

Click a document name to see the full chunk breakdown:

• Each chunk's text preview
• Token count per chunk
• Embedding status

This is useful for debugging why a specific question isn't being answered correctly.

Best Practices

Follow these tips to get the most out of your knowledge base:

1. Upload FAQs first — FAQs have the highest ROI because they directly map customer questions to answers. If you only upload one document, make it your FAQ.

2. Keep documents up to date — Outdated information is worse than no information. Set a reminder to review your knowledge base quarterly.

3. Use clear headings — The chunker uses headings as natural split points. Well-structured documents (with ## Heading 2 or ### Heading 3 in Markdown, or proper heading styles in Word) chunk more cleanly.

4. Test with real customer questions — Don't just upload and assume it works. Test with actual questions customers have asked in the past.

5. Start small, expand as needed — Begin with 3-5 core documents (FAQ, pricing, product overview). Monitor AI conversations to spot gaps, then upload targeted documents to fill those gaps.

6. Avoid redundancy — Don't upload 10 versions of the same FAQ. Redundant chunks dilute search relevance.

Troubleshooting

AI Gives Generic Answers

Problem: The AI responds with "I don't have that information" or gives a generic ChatGPT-style answer instead of using your knowledge base.

Possible causes:

• Document is still processing (check the status badge)
• The question isn't covered in your uploaded documents
• The question is phrased very differently from the content in your docs

Solution: Upload more comprehensive documentation or rephrase the content in your docs to match common customer phrasing.

Upload Fails

Problem: Upload button spins indefinitely or shows an error.

Possible causes:

• File is too large (over 10MB)
• Unsupported file format (e.g., scanned PDF without OCR)
• Network timeout

Solution: Check the file size and format. For large PDFs, try splitting into smaller files. For scanned PDFs, use OCR software to convert to text first.

Wrong or Irrelevant Answers

Problem: The AI gives an answer from the knowledge base, but it's not the right answer for the question.

Possible causes:

• Document content is ambiguous or contradictory
• Multiple documents have similar-sounding but different information
• The chunk retrieved was missing important context

Solution: Review the document in the knowledge base UI. Add more specific FAQ entries that directly match customer questions. Use clear language and avoid jargon.

What's Next?

Now that you have an AI knowledge base powering your RAG WhatsApp auto-replies, explore these related features:

• Train AI Style Profiles — Customize the tone, personality, and response style of your AI auto-reply
• Build Support Bot with OpenClaw — Combine knowledge base with AI agents for advanced automation
• REST API Quick Start — Upload documents programmatically via the API
• Send Your First Message — Get started with basic WhatsApp automation

Your AI knowledge base is the foundation of intelligent automation. Keep it fresh, test it regularly, and watch your AI auto-reply quality improve over time.

Build your first automation in 5 minutes — start your free trial

Need help? Contact support via the dashboard chat or email us at [email protected]. We typically respond within 2-4 hours.