Semantic search

Introducing Semantic Search with LLM Summarisation and a guide for users and editors. This innovative technology will allow you to get more accurate and contextually relevant search results. In this page, we will explain how to use this feature effectively as well as offer tips to our editors on optimising their content for best results.

Basic notions

Toolboxes usually have large content in various formats, including toolbox pages and subpages, attachment lists, and various attached files like PDF and Word documents.

The basic search function finds pages and PDF documents that contain the typed keywords, and links to them. It cannot search for synonyms or variants of the keywords.

The semantic search function determines the notions appearing in the search query, and performs a similarity comparison within the content of all pages and documents on the toolbox. The most similar document pieces are then sent to a Large Language Model (LLM) which summarises the relevant information. The relevant documents should appear linked, so that a user can open and review them.

Like the basic search you can type keywords, but the semantic search should understand questions, contexts, synonyms and variants of the keywords. Unlike the basic search, you can type a natural language question (in English).

This is experimental technology, and the quality of the answers depend on the availability and quality of information that editors have added to the toolbox, and the answers may be incomplete or wrong. Users must open and review the linked documents instead of relying on the summary.

Note that while we use LLMs like some artificial intelligence (AI) chat bots, the semantic search is not conversational. Your queries are unique, the LLM doesn't keep track of your previous questions. You can see your previous questions and results pushed to the bottom of the search results in collapsed sections.

For users

Access to the semantic search function is currently in testing, limited to the website, and to editors and administrators. To try it, go to "Wrench menu -> Administration panel" then near the top click on the link "Try the experimental LLM semantic search".

Type search keywords or a question into the search box and press "Enter". After a few seconds, the results will appear below. You will see summary answers from the LLMs, and a list of document links that were estimated to most likely answer your query.

You can review the answers and the linked documents. Remember, the semantic search results depend on the quality of the information added by your colleagues from the trust-based editorial team, and answers might be imprecise, or it may lack information, or occasionally present niche information as general. Users are responsible of checking the actual documents and using the information from them with their professional judgement.

If the initial answers and results don't seem relevant to your query, don't hesitate to refine your search terms and try again. Sometimes, a slight adjustment in keywords or rephrasing can lead to more accurate results.

You can optionally rate the answers and send feedback to the editing team which would then have the opportunity to improve the underlining knowledge base by adding or updating the relevant documents.

For editors and administrators

The semantic search depends on the knowledge base that is added to the toolbox.

You can access previous search queries and results from the link "Review search history" in the top section of the page. You will see previous questions and answers, and possibly user rating and feedback. The most recent queries will be at the top, and you can delete individual results.

If an answer is not optimal, verify that the information is there, if not, add it. Don't hesitate to place relevant keywords in different sections of the document or toolbox page, this should increase the results precision. After saving a page, or uploading a new file, you can search again to see the new results. (Note, updated pages and new files are instantly added to the search index, but detached or replaced files are removed only the next day.)

In your toolbox pages and attached files:

Employ clear headings and subheadings: structuring your text with descriptive section titles improves comprehension and assists the semantic engine in identifying pertinent sections. Use specific terminology reflecting your content's focus.
Include relevant keywords: add appropriate keywords and notions describing each page's topic, ensuring efficient indexing by our semantic database. These should include for example both generic and commercial medication names, and both popular and scientific disease names written in the document.
Separate the text into paragraphs with double line breaks between them. The paragraphs most similar to the query will be presented to the LLM as context (knowledge source) and to the user as search results.
Focus on quality writing: ensure clarity, coherence, and relevance throughout your texts. Highlight key concepts early on and repeat crucial details periodically to strengthen the semantic discoverability and enhance overall readability.
Optimise file names: name attachments and toolbox pages meaningfully, incorporating essential keywords related to contents therein.
Revise frequently: regularly update existing content based on feedback, analytics data, and emerging trends. Refreshing material not only boosts engagement but also encourages continuous improvement in our semantic search algorithms.

Note that:

There is a short delay between updating a page or uploading a file, and its indexing and availability in the semantic search.
Attached pictures in pages, or pictures inside PDF or office documents, are not indexed, that is, if there is text on a picture, this text will not be available in the semantic knowledge base. This is the case for some charts, or scanned documents.
Only toolbox pages and attachments of the file formats PDF, DOCX, DOC, and ODT are indexed and searched. If you have other file formats such as PPTX, please export the file to PDF and upload that one instead.
Attached files need to NOT be password-protected in order to be discoverable.

In conclusion, the Semantic search with LLM summarisation will eventually improve the browsing experience by delivering precise, valuable, and easily accessible insights tailored to individual needs. You can contribute significantly toward enhancing this tool while fostering a vibrant community centred around shared knowledge and discovery.