Use Human Feedback

Relevance is in the eye of the beholder: a certain document may be relevant for a query in one domain, but irrelevant in another. Consider the following notional document and query:

• Query: health benefits of exercises
• Document Title: The Benefits of Yoga for Stress Reduction and Mental Health

In the domain of wellness and mental health, the document title and search query would be a good match; they both refer to the benefits of exercise on mental health.

However, in the domain of sports medicine or physical therapy, the document would not be an appropriate match for the query, as the query could be interpreted as looking for information on the physical benefits of exercise, which is unrelated to the mental health benefits of yoga mentioned in the document.

Examples like these are particularly problematic for keyword-based search engines and less so for Noematic, but we can do even better by incorporating human feedback. Through positive and negative feedback signals, Noematic can automatically adapt its search strategy to your unique domain or operating environment.

Sending feedback signals

Suppose you're an HR officer and searched for best strategies for reducing stress at work and your search surfaced the following top 2 results:

1. How to Eliminate Workplace Stress by Changing Jobs: A Comprehensive Guide
2. 10 Proven Techniques to Reduce Stress and Increase Productivity in the Workplace

In your role of HR officer, the second result might be a good match for your query, but the first may not. You can use the create_database_feedback action to better align future searches with your intentions.

{
  "version": "1.0",
  "view": "team",
  "action": "create_database_feedback",
  "data": {
    "database": {
      "id": "MtjraphUEnYS8esgGgTS12Q7zs4uyBrHw"
    },
    "query": "best strategies for reducing stress at work",
    "content": "How to Eliminate Workplace Stress by Changing Jobs: A Comprehensive Guide",
    "signal": -1
  }
}

{
  "version": "1.0",
  "view": "team",
  "action": "create_database_feedback",
  "data": {
    "database": {
      "id": "MtjraphUEnYS8esgGgTS12Q7zs4uyBrHw"
    },
    "query": "best strategies for reducing stress at work",
    "content": "10 Proven Techniques to Reduce Stress and Increase Productivity in the Workplace",
    "signal": 1
  }
}

This action takes the following parameters:

• objectdatabase
  is the database to align based on this feedback.
  • stringid
    is the globally unique identifier of the selected database.
• stringquery
  is the search query pertaining to the feedback.
• stringpassage
  is an excerpt from a search result that you'd like to provide feedback for.
• floatsignal
  is a value between -1 (extremely irrelevant) and 1 (extremely relevant).

Collecting human feedback from end users

While it's entirely possible to to curate your feedback internally, it's often desirable to allow your end users to contribute to the process. This can be readily achieved using "thumbs up" and "thumbs down" buttons placed near every search result. When you receive such an indications by users, you can map them to signal scores of 1 and -1 respectively and proxy them to Noematic.