What Are Embeddings and How Do Chatbots Use Them

Embeddings in Plain Language

Think of embeddings as a way to turn words into coordinates on a map of meaning. Two sentences about the same topic end up close together on this map, even if they use completely different words. "What is your return policy?" and "Can I send back an item I purchased?" have very different words but nearly identical meaning, so their embeddings are close together.

This is what makes chatbot knowledge search so powerful. A customer does not need to use the exact keywords from your documentation. They can phrase their question however they want, and the embedding search finds the most relevant content because it matches meaning, not keywords. This is a major improvement over older keyword-based search systems.

How Embeddings Are Created

When you upload a document or crawl your website, the system breaks the content into chunks of roughly 200 to 500 words. Each chunk is sent to an embedding model (a specialized AI model designed just for this purpose) which returns a long list of numbers representing the text's meaning. These numbers are stored alongside the original text in the embeddings database.

On this platform, embedding creation costs 3 credits per chunk. This is a one-time cost when the content is uploaded. The embeddings persist in the database and can be searched unlimited times at no additional cost per search. The search cost is included in the chatbot's per-message pricing.

How Search Works at Query Time

When a visitor sends a message to your chatbot, the following happens behind the scenes:

1. The visitor's message is converted into an embedding using the same model that created the document embeddings
2. The system compares this query embedding against all stored embeddings for the chatbot's knowledge base
3. The top matching chunks (typically 3 to 5) are retrieved based on similarity scores
4. These chunks are included in the prompt sent to the AI model (GPT, Claude, etc.) along with the conversation history and system instructions
5. The AI model reads the retrieved chunks and generates a response that draws from that specific content

This entire process happens in milliseconds. The visitor experiences a natural conversation with no visible delay from the search step. See What Is RAG and How Chatbots Use It for the full pipeline explanation.

Why Chunk Size Matters

The chunk size directly affects retrieval quality. Chunks that are too small (a single sentence) may lack enough context for the AI to generate a useful answer. Chunks that are too large (an entire page) may contain too much unrelated information, diluting the relevant content with noise.

The sweet spot is a complete thought or a focused section, usually 200 to 500 words. A chunk about your return policy should contain the entire return policy, not just the first paragraph. A chunk about product specifications should cover one product, not your entire catalog. The platform handles chunking automatically, but you can improve results by structuring your documents with clear section breaks. See How to Chunk Documents for Better AI Understanding.

Embeddings vs Keywords

Traditional search (like a search bar on your website) matches exact keywords. If your document says "refund" but the customer types "money back," keyword search fails. Embedding search succeeds because both phrases have similar meaning and therefore similar embeddings.

This semantic understanding is especially valuable for chatbots because customers phrase questions in unpredictable ways. They might ask "how long till it ships," "when will I get my order," or "delivery timeline." All three map to the same embedding neighborhood as your shipping policy document, so the chatbot finds the right content regardless of wording.

Managing Your Embeddings

Embeddings are organized by tags in your admin panel. You can view how many chunks exist for each tag, delete embeddings for outdated content, and add new embeddings as your content grows. When policies or product information change, delete the old embeddings for that topic and upload the updated document to create fresh ones. The chatbot immediately starts using the new content.