Knowledge bases

Knowledge bases are collections of documents that your copilot uses to answer user queries. They are crucial for ensuring that your copilot performs well and provides accurate, domain-specific information. Typical knowledge bases include:

Product documentation
Marketing materials
Support knowledge bases
Real-time user and application data

Continual generates a semantic index of knowledge base documents, allowing copilots to quickly retrieve relevant information before responding to a user.

Overview

Continual supports two types of knowledge bases:

Managed knowledge bases

Continual automatically updates documents pulled from external sources like websites or databases every hour, day, or week. You can also sync the knowledge base manually.

Supported managed data sources:

Website - Enter links to product documentation, public knowledge bases, blogs, or other web-based information sources.
Cloud Blob Storage - Sync a cloud bucket with a Continual knowledge base. Supported providers:
- Azure Blob Storage (opens in a new tab)
- AWS S3 (opens in a new tab)

Unmanaged knowledge bases

Documents are added, updated, and deleted freely without any source constraints. Continual automatically indexes the provided files but does not sync them with an external source.

Supported unmanaged data sources:

Files - Upload PDF, text, CSV, YAML, or JSON files. Files are converted into Markdown (opens in a new tab) format and can be edited in the Continual console.

Excel/CSV Spreadsheets - Each row maps to a single document in Continual:

Column	Required?	Description
Name	Yes	Document title
Content	Yes	Document content
PublicURL	No	Allows copilot to provide public reference links in answers
PrivateURL	No	Allows copilot to provide private reference links in answers
Metadata1...N	No	Additional metadata columns (author, dates, content owner, etc.)

Improving performance

There are several ways to improve the performance of your knowledge base.

Monitoring

Continual provides insights into how your knowledge base is being used, such as:

Statistics on which copilots are using the knowledge base
Query relevance, speed, and frequently referenced documents
Insights for content managers to learn what users want to know

Best Practices

Create an FAQ knowledge base

Many times, your existing product documentation or knowledge bases may not cover questions that users often ask. To address this, you can create a knowledge base dedicated to "Frequently Asked Questions."

For optimal query performance, store each question-answer pair as a separate document with the question as the document name and the answer as the document content. This helps the copilot easily find potential answers to user questions and also allows you to easily see which questions are commonly referenced.

As you move into production, you can identify unanswered questions using our monitoring tools and add them to the FAQ knowledge base.

Index both application and user data

To provide timely and precise responses to user queries, you should consider indexing real-time application and user data, instead of only providing it through a tool. This allows the copilot to more efficiently find and combine information.

Make sure to employ metadata filtering to ensure that users only see the information they are authorized to access.

Include specialized, domain-specific knowledge bases

The value your copilot offers increases when you include specialized data that's unique to your domain or product.

For instance, if your application serves the real estate market, you might consider adding a knowledge base filled with information about the various properties listed on your platform.

Metadata filtering

Metadata filters restrict document access based on user, group, team, department, company, or context. The copilot can only access documents relevant to what the user is currently viewing.

By default, knowledge bases are always included in copilot retrievals but can be filtered by a signed JWT grant or client filter if a matching metadata field is provided. Private knowledge bases require a signed JWT grant to be retrievable.

Applying metadata to documents

Set metadata automatically on upload/ingestion or manually in the UI:

Manual - Edit document, set metadata key and value, save
Automatic
- Cloud Bucket: Sync tags from buckets, folders, and files
- CSV: Additional columns after Name and Content are added as metadata
- Website: Capture metadata fields (title, author, keywords, etc.) from webpage
- Files: No automatic metadata

Applying metadata filters from a copilot

Set filters in your copilot widget client-side or as a JWT grant.

Client-side - Ad-hoc filters based on user's current application location

<CopilotProvider datasetFilters={{"mydatasetid": {metadata_key: metadata_value}}}>
{...}
</CopilotProvider>

JWT Grant - Securely filter documents by user or group

const generateUserAccessToken = (userId, copilotId, copilotSecret) => {
  const algorithm = "HS256";
  const ttl = 86400;
  const payload = {
    grants: {
      identity: userId,
      datasets: { mydatasetid: { my_metadata_key: my_metadata_value } },
    },
  };
  const header = {
    typ: "JWT",
    alg: algorithm,
  };
  const signOptions = {
    header: header,
    issuer: copilotId,
    expiresIn: ttl,
  };
  console.log(typeof copilotSecret);
  console.log(typeof signOptions);
  return jwt.sign(payload, copilotSecret, signOptions);
};

Testing metadata filters

Test metadata filters in the Continual Console Playground:

Set metadata on connected knowledge bases.
Set client filters by clicking Impersonate.

Copilots Tools