Template Query

POST

api

collection

template

curl --request POST \
  --url https://app.twilix.io/api/v1/collection/template \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{
  "collection": "<string>",
  "prompt": "<string>",
  "promptFields": [
    {}
  ],
  "fields": [
    {}
  ],
  "conversationID": "<string>",
  "minimumScore": 123,
  "limit": 123,
  "includeModeration": true,
  "stream": true
}'

For users wanting more flexible control over prompting and outputs as part of their RAG solution, they can do this with our template endpoint.

You want to use this endpoint if you are looking for:

More fine-grained control over what the model outputs such as specifically HTML or Markdown.
More steerable inputs where you want to provide an example response before adding references into the prompt.

collection

string

required

The collection to query.

prompt

string

required

The template that you want to use. This template uses a reference magic in order to provide users with more flexible control over their LLM outputs.

An example template that is looking for just the returned Markdown could be:

You are a cybersecurity consultant, can users help provide
a clearer understanding of what is happening? Return 
this in Markdown with clear headings to separate it out.
{reference}
Markdown:

On our backend, we replace reference with the relevant promptFields that you supply. If None is supplied, then it uses all fields in a collection.

promptFields

array

The fields that you want to use to feed into the prompt template.

fields

array

The fields that you want to be returned as reference. If not specified, it returns all fields as reference.

conversationID

string

The conversation ID. This is returned in the response so you can use the one that has been automatically generated for you or you can also supply your own to keep track of the conversation on your side.

minimumScore

float

The minimum rerank score.

limit

integer

The max number of documents to returned

includeModeration

boolean

If true then there will be a moderation layer applied after the user inputs a query and when the AI outputs to ensure that the generated content is not harmful or violent.

stream

boolean

Whether or not a stream response should be returned. See examples below for details.

Content-Type

default: "application/json"required

Requires JSON Content Type

Since it could take up to a few seconds for the answer to complete. Streaming is encouraged for the best user experience.

You can pass in stream=true in the request to enable streaming response, and easily setup stream handling with Microsoft’s fetch-event-source package.

Example.ts

const requestBody = {
  collection: "insurancePdfs",
  prompt: "Can I claim wear and tear?",
  promptFields: ["content", "url", "pageNumber"],
  stream: true, // <------- IMPORTANT
}; // Define your request body

const options = {
  method: "POST",
  headers: {
    Authorization: `Bearer ${apiKey}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify(requestBody),
  signal: abortSignal, // This is optional
};

await fetchEventSource(`https://app.twilix.io/api/v1/collection/copilot`, {
  ...options,
  onopen(response) {
    // ...
  },
  onmessage: (event) => {
    const decodedJson = JSON.parse(event.data);
    const decodedMessage: GenerativeQnAResponse = {
      message: decodedJson.message,
      conversationID: decodedJson.conversationID,
      references: decodedJson.references ?? [],
      confidenceScore: decodedJson.confidenceScore,
    };
    onNewData(decodedMessage);
  },
  openWhenHidden: true,
  onerror: (error) => {
    // ...
  },
});

The event stream looks like this:

data: {"message": ""}
data: {"message": "Hello"}
data: {"message": "!"}
data: {"message": " How"}
data: {"message": " can"}
data: {"message": " I"}
data: {"message": " assist"}
data: {"message": " you"}
data: {"message": " today"}
data: {"message": "?"}
data: {"message": "", "conversationID": "f4eb4cdf-bf94-438f-9143-8d96d6d4d0d8", "references": [{"reference": "Dummy PDF file\n", "content": "Dummy PDF file\n", "url": "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf", "pageNumber": 1}], "confidenceScore": null}

If latency is not a concern, you can alternatively use the normal calling method and wait for the entire response.

Note it could take up to a few seconds for the answer to complete.

Example.ts

const requestBody = {
  collection: "insurancePdfs",
  prompt: "hi",
  stream: true, // <------- IMPORTANT
}; // Define your request body

return axios
    .post(`https://app.twilix.io/api/v1/collection/copilot`, requestBody, {
        headers: {
            Authorization: `Bearer ${apiKey}`,
            "Content-Type": "application/json",
        },
    })
    .then((res) => {
        // Handle Response here
        return res.data;
    })
    .catch((err) => {
        // Handle Error
    });

The response looks like this:

{
  "message": "Hello! How can I assist you today?",
  "conversationID": "01f02c68-584a-42d0-bc1c-71016b127f93",
  "references": [],
  "confidenceScore": 1
}

Question Answering vs Copilot Schema

curl --request POST \
  --url https://app.twilix.io/api/v1/collection/template \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{
  "collection": "<string>",
  "prompt": "<string>",
  "promptFields": [
    {}
  ],
  "fields": [
    {}
  ],
  "conversationID": "<string>",
  "minimumScore": 123,
  "limit": 123,
  "includeModeration": true,
  "stream": true
}'

Introduction

Ingest Data

Query Data

Vectorstore

Tutorials

Blog