Knowledge Base QA and Search

How to use knowledge base QA and search in the Platform

Introduction

qa.search and qa.answer functions access the knowledge base to answer user questions. Additionally, we provide information about the KBDocument class, which is essential for understanding the data structure used by qa.search.

Table of Contents

1. Using qa.search

2. Using qa.answer

3. Understanding the KBDocument Class

---

1. Using qa.search

The qa.search function is used to search for information within the knowledge base. It allows you to retrieve relevant documents that contain potential answers to your questions.

How to Use qa.search

val results = qa.search(question, numberOfResults)

• question (String): Your question or query.

• numberOfResults (Optional, default=3): The number of relevant documents to retrieve.

Returns

• results (List of KBDocument objects): A list of relevant documents containing information related to the query.

Example

Suppose you want to find information about "climate change". You can use qa.search like this:

val searchResults = qa.search("What is climate change?")

This will return a list of relevant documents that may contain answers to your question.

---

2. Using qa.answer

The qa.answer function is used to generate answers to specific questions based on the retrieved documents from the knowledge base via a large language model.

How to Use qa.answer

val answer = qa.answer(question, prompt)

• question (String): Your question or query.

• prompt (String): Additional context or information to assist in generating an answer.

Returns

• answer (String): The generated answer to the question based on the retrieved documents and provided prompt.

Example

Suppose you want to generate an answer to the question "What are the causes of climate change?". You can use qa.answer like this:

val answer = qa.answer(
    "What are the causes of climate change?",
    "Please answer the question if it is answerable based on the knowledge provided."
)

In the first step, qa.answer utilizes the qa.search function to retrieve relevant documents from the database, in the second step, it feeds the documents together with your question and prompt to a large language model in order to provide an answer. This will generate an answer based on the retrieved documents and the generic prompt provided. The full prompt and logic can be seen below.

Source Code for qa.answer Function:

fun qa.answer(question: String, prompt: String): String {
    val retrievedDocuments = qa.search(question)
    val combinedPrompt = "$prompt\nKnowledge:\n${retrievedDocuments.joinToString(separator = "\n", prefix = "- ")}\nQuestion: $question\nAnswer:"
    val answer = largeLanguageModel.complete(combinedPrompt)
    return answer
}

---

3. Understanding the KBDocument Class

The KBDocument class is a data structure used within the Platform to represent information retrieved from the knowledge base. It contains the following properties:

• answer (String): This field is empty in the current knowledge base configuration, as the answer generation is handled by an LLM.

• score (Double): A numerical value representing the relevance or confidence level of the document.

• context (String): The full retrieved document in string form.

• documentIds (List of Strings): Unique identifiers for the document.

Example

Here is an example of a KBDocument object:

val document = KBDocument(
    answer = "",
    score = 0.85,
    context = "Climate change refers to long-term shifts in temperatures and weather patterns. These shifts may be natural, but since the 1800s, human activities have been the main driver of climate change, primarily due to the burning of fossil fuels (like coal, oil, and gas) which produces heat-trapping gases.",
    documentIds = listOf("d2a0b178c8e6cda845b86b203fd2531f")
)

In this example, the KBDocument represents information about the causes of climate change. You can access different fields of the KBDocument class in Kotlin as follows:

val answerText = document.answer
val relevanceScore = document.score
val additionalContext = document.context
val uniqueIds = document.documentIds

These variables will contain the respective values from the KBDocument object.