Quickstart: your first cited answer

Getting a first cited answer takes four things: a collection, a source synced into it, an assistant over that collection, and a question. The six steps below take an empty organisation to a grounded, cited answer. They use the HTTP API throughout, so each step can be copied as-is, and the console equivalent is noted beside each. To put an assistant on your website without writing any code, use the console’s three-step website setup instead.

Before you begin

You need:

A MITHUNAI organisation. Ask for one with Request access on the sign-in screen (a person approves each request; see Request access), or accept a colleague’s invitation. There is no instant sign-up. See Onboarding.
An API key with the Editor role or higher. An Owner or Administrator creates it in the console under API keys. See API keys.
Content to answer from: a documentation site, sitemap, GitHub repository or set of files. Connectors lists every source MITHUNAI reads.

Each answer an assistant gives counts against your organisation’s monthly answer allowance, wherever it is asked. Set two environment variables. Every example below uses them.

export MITHUNAI_URL="https://app.mithunai.com"  # your MITHUNAI host
export MITHUNAI_API_KEY="mithunai_sk_..."          # the key you created

An API key is a secret. Keep it on servers and in secret managers. Never put it in a web page, a mobile app or a repository.

Confirm a model is available

curl "$MITHUNAI_URL/mithunai/api/v1/models" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY"

Pick a model whose capabilities include chat and save its id, such as openai/gpt-4o, as MODEL_ID. If the list is empty, no model provider is configured: contact whoever operates your MITHUNAI deployment.

Create a collection

A collection is the corpus an assistant answers from. Its embedding model is fixed when you create it.

curl --request POST "$MITHUNAI_URL/mithunai/api/v1/knowledge/collections" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{"name": "Product documentation"}'

Save the id from the response as COLLECTION_ID.

In the console, the Sources page creates a collection for you the first time you connect a source.

Connect a knowledge source

Register your documentation site. A sitemap is usually faster and more complete than a crawl; see Knowledge sources for every source type.

curl --request POST "$MITHUNAI_URL/mithunai/api/v1/knowledge/sources" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "collection_id": "'"$COLLECTION_ID"'",
    "source_type": "sitemap",
    "name": "Docs site",
    "configuration": { "seed_url": "https://docs.example.com/sitemap.xml" }
  }'

Save the source id as SOURCE_ID. In the console: Sources → Connect source.

Ingest it

Registering a source does not read it. Start an ingestion job, then poll it until it finishes.

curl --request POST "$MITHUNAI_URL/mithunai/api/v1/knowledge/sources/$SOURCE_ID/sync" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY"

curl "$MITHUNAI_URL/mithunai/api/v1/knowledge/sources/$SOURCE_ID/job" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY"

Ingestion runs in the background and a large site takes minutes. Wait for the job to complete before asking questions. See Ingestion.

Create an assistant over the collection

curl --request POST "$MITHUNAI_URL/mithunai/api/v1/assistants" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "name": "Docs assistant",
    "model": "'"$MODEL_ID"'",
    "knowledge_base_ids": ["'"$COLLECTION_ID"'"]
  }'

Save the assistant id as ASSISTANT_ID. Without model, the deployment picks a chat model it can reach, preferring anthropic/claude-sonnet-5. A model you name must be one your deployment offers, or the request is refused with 400.In the console: Customizations → New assistant, then tick the collection under Knowledge. See Create and configure assistants.

An assistant answers only from the knowledge attached to it, and refuses questions until at least one collection is attached.

Ask a question

Start a conversation, then ask.

CONVERSATION_ID=$(curl --silent --request POST "$MITHUNAI_URL/mithunai/api/v1/conversations" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{"assistant_id": "'"$ASSISTANT_ID"'"}' | jq -r .id)

curl --request POST "$MITHUNAI_URL/mithunai/api/v1/conversations/$CONVERSATION_ID/messages" \
  --header "Authorization: Bearer $MITHUNAI_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{"text": "How do I get started?"}'

The response contains the answer and its citations. If the answer has "abstained": true, your content did not support an answer. That is a correct result, not an error. See How answers work.You can also ask in the console in the Playground.

Next steps

Embed the widget

Put the assistant on your website by pasting one line.

Stream answers

Show the answer as it is written.

Test before launch

Check grounded answers and abstentions against your own questions.

Connect MCP clients

Query your knowledge from Claude and IDE assistants.

Compare the channels

Which of the widget, the API and MCP suits who is asking.

When a step goes wrong

Empty model lists, failed ingestion jobs, and what a 401 or 429 means.

Get started

Resources

Quickstart: your first cited answer

Before you begin

Next steps

Embed the widget

Stream answers

Test before launch

Connect MCP clients

Compare the channels

When a step goes wrong

​Before you begin

​Next steps

Embed the widget

Stream answers

Test before launch

Connect MCP clients

Compare the channels

When a step goes wrong

Before you begin

Next steps