Functionality

This module is designed to answer a user query based on web pages retrieved using a web search engine. The module behaves as a standard GPT but has access to a web search tool. When the user asks a question that the model doesn’t know the answer, it executes the following workflow:

1. Formulates a search query optimized for search engines (based on the entire discussion history)

2. Executes a web search using the configured search engine.

3. Scraps the content of top web pages returned by the search engine

4. Divide the scrapped content into small chunks of text

   1. [Optional] Embed the chunks locally and perform a similarity search based on the generated query.

5. Passes the relevant information to the context and answers the user’s question/query

Input

A user question

Example input:

• “What are the recent trends and developments in Zurich’s financial and insurance markets?“

Output

An answer based on retrieved web pages, either referencing the appropriate links or stating that no information was found in the retrieved web pages.

Configuration settings (technical)

Web Search Config

Provided below is the general structure of the configuration for the Web Search module.

Bing Search API

Bing Custom Search Config

Bing offers the possibility to customize the search results. For example:

• Limit the search to a predefined set of domains

• Promote a domain

• Pin results from a specific website

• Filter links

• …

This Web Search module developed by Unique supports custom search through the custom_config_id parameter that can be obtained from https://www.customsearch.ai/. More details on setup can be found in the following link: https://learn.microsoft.com/en-us/bing/search-apis/bing-custom-search/how-to/quick-start

Once the id is created, it can be passed to through config to your assistant.

There is also the possibility to boost specific search domains without the need to create a custom search id. This can simply be achieved by passing the list of domains in the boosted_search_domains parameter. NOTE: contrary to custom search, this method doesn’t necessarily restrict the search domains to the list provided in boosted_search_domains.

Google Search API

Google Custom Search Config

Google Custom search is not supported yet, but can be implemented if requested by users.

Data Source Config

This configuration defines how the module deals with the data retrieved from the internet (e.g. number of sources, chunk_size, …)

Answer Generation Config

This configuration defines how the module make use of the retrieved pages from the internet to generate its answer.

Query Generation Config

The module automatically optimize the question of the user to better work with search engines. The user for example may want to always search the web in English or as in the default value, use the

Example of configuration

{
  "WebSearchConfig": {
    "model_name": "AZURE_GPT_4_0613",
    "search_engine_name": "bing",
    "bing_config": {
      "search_country_code": "CH",
      "search_market": "de-CH",
      "custom_search_config": {
        "id": "your_custom_config_id",
        "boosted_search_domains": []
      }
    },
    "data_sources_config": {
      "fetch_size": 4,
      "max_workers": 10,
      "chunk_size": 600,
      "chunk_overlap": 50,
      "embedding_batch_size": 128
    },
    "answer_generation_config": {
      "limit_token_sources": 5000,
      "max_chunks_to_consider": 20,
      "number_history_interactions_included": 2,
      "embedding_reranking": true
    },
    "query_generation_config": {
      "query_instruction": "The user's search query, optimized for search engines, incorporates relevant details from the conversation, especially if it is a follow-up question. Always use the user's language message."
    },
    "debugging": false
  }
}