Knowledge bases

Answers grounded in your documents, not the model's imagination

Upload your pricing sheets, policies, FAQs and product guides. ReplySetter chunks and embeds them into the vector store you choose, and agents search it with a tool whenever a question needs facts, so replies quote your truth.

A policy document is chunked, embedded into vectors and indexed; a customer question finds the nearest passages and the agent answers with a citationReturns.pdf#1#2#3#4[0.12, -0.44, 0.08…][0.31, 0.02, -0.17…][-0.05, 0.27, 0.66…][0.44, -0.12, 0.29…]embedded · 1536 dims (auto-detected)VECTOR INDEXqueryCan I return something Ibought in the sale?Yes! Sale items can be returned within14 days for store credit or an exchange.Returns #3
4vector store options
0setup for the built-in store
Autoembedding dimension detection
Per stepknowledge access
Deep dive

A closer look at Knowledge bases

01 · Ingest

Chunked, embedded and indexed automatically

Add a document and it's split into overlapping chunks, embedded with the model you choose, and written to the store. Embeddings come from OpenAI, OpenRouter or Cloudflare Workers AI, and dimensions are detected on first use, so there's nothing to configure.

  • Automatic chunking and embedding
  • OpenAI, OpenRouter or Workers AI embeddings
  • Dimension detection on first use
  • Re-index a document or a whole base in one click
A document is cut into overlapping chunks, embedded by the chosen model and written to the vector tablePricing guide.docxchunk 1chunk 2chunk 3overlap keeps context across cutsEmbedding3-small · 1536dkb_supportid 1041vector(1536)id 1042vector(1536)id 1043vector(1536)id 1044vector(1536)indexed
02 · Stores

Bring the vector store you already trust

Start with the built-in pgvector store in the app's own Postgres, or point a knowledge base at your own pgvector or Supabase database, a Pinecone index or Cloudflare Vectorize. Each store keeps knowledge bases cleanly separated.

  • Built-in pgvector: nothing to set up
  • pgvector / Supabase: one table per knowledge base
  • Pinecone: one namespace per knowledge base
  • Cloudflare Vectorize: separated by metadata
One knowledge base can live in built-in pgvector, Supabase, Pinecone or Cloudflare VectorizeSupport KB86 documents · 2,340 chunksBuilt-in pgvectorIn the app's own Postgres. Zero setup.syncingpgvector / SupabaseYour Postgres, one table per base.syncingPineconeOne index, a namespace per base.syncingCloudflare VectorizeOne index, bases split by metadata.syncing
03 · Retrieve

The agent searches when it needs to, and only then

Retrieval is a tool, not a blob pasted into every prompt. The agent calls search_knowledge with a focused query, reads the best-matching passages and answers from them. You can see every query and result in the debugger.

  • search_knowledge is called on demand
  • Focused queries, relevant passages
  • Every search visible in the debugger
  • Diagnose can suggest knowledge fixes
The agent calls search_knowledge, ranks passages by similarity and answers from the top matchesDo you ship to Canada?Tool call · search_knowledgesearch_knowledge({ query: "shipping to Canada, duties" })Shipping.mdWe ship to Canada via DHL in 5–7 days…0.89Duties.mdDuties and taxes are included for CA…0.81Returns.pdfInternational returns are accepted…0.62Yes! We ship to Canada with DHL in 5–7 businessdays, and duties and taxes are already included.query + results in the debugger
Under the hood

The details

Built-in

pgvector in the app's own Postgres. No setup required.

pgvector / Supabase

Any Postgres with pgvector; one table per knowledge base.

Pinecone

One index, with a namespace per knowledge base.

Cloudflare Vectorize

One index; knowledge bases told apart by a metadata field.

Embeddings

text-embedding-3-small, bge-m3 and more via OpenAI, OpenRouter or Workers AI.

Search API

Search any knowledge base from the app or over MCP to check what agents will find.

Use cases

Who it's for

Support

Warranty, returns and shipping answers straight from your policy docs.

Sales

Pricing tiers and package details quoted accurately every time.

Onboarding

Product guides that answer “how do I…” questions over chat and email.

FAQ

Knowledge bases: common questions

Do I need my own vector database?

No. The built-in store uses pgvector inside the app's own Postgres with zero setup.

Can I use Pinecone or Supabase?

Yes, along with any pgvector database and Cloudflare Vectorize.

Is my whole knowledge base sent with every message?

No. The agent searches with a tool and only reads the passages it needs.

How do I check what the agent found?

Open the debugger on any reply to see each knowledge search and its results.

Get started

Put it to work on your own inbox

Connect an inbox, describe the job, test it in a sandbox and go live with exactly the autonomy you're comfortable with.