The objective is to create a “Chat-GPT like” user experience on documents, which are often internal and confidential: ask questions to documents and obtain answers that contain a link to the source document.
- Create a project and select the Question Answering option.
Document upload and conversion
- Upload and convert documents automatically. A few documents are enough to test the solution.
- It is possible to customize document upload and convert from formats such as audio (speech to text), PDF-image or scientific literature.
Document segmentation (chunking)
- Documents are then automatically segmented in so-called ‘chunks’ or ‘snippets’, typically a phrase or paragraph.
- It is possible to customize document segmentation by using off-the-shelf segmenter or by creating a segmentation pipeline.
- While for plain text search the uploaded documents are indexed, for semantic search (similarities) the segments are automatically vectorized, a process that consists of creating numeric values from the words with so-called embeddings.
- It is possible to select other off-the-shelf embedding models such as OpenAI embeddings or create custom embedding models
Ask a question – Search
Three different types of search exist:
- Full text search matches the text elements from the question to relevant elements in snippets
- Semantic search uses embedding-based vectors to find similarities
- Hybrid search is a combination of both methods
- The search results (the corresponding text snippets) are displayed and ordered by relevance. The technical process used to generate these search results is called a Retriever.
- For semantic and hybrid search it is possible to select different embedding models (Vectorizers) if these are configured.
With the help of professional services it is also possible to create custom vectorizers, for exemple focusing on a specific language or business domain.
Ask a question – answer generation
- Once document have been successfully imported, automatically indexed and vectorized, go to the Question-answering menu
- Enter your question in the search box and press “Enter”
- A search strategy, a vectorizer and a LLM are used by default.
- Large Language Models (LLMs) are used to generate an answer from the search results. The answer is presented with links to the relevant snippet, that in turn contains a link to the source document.
- Different LLMs can be configured and selected.
- Commercial LLMs such as the ones provided by Open AI are subject to costs and are integrated using an API. These costs are minimized by only sending the snippets that are the results of a search to the LLM.
- Private LLMs such as Llama2, Mistral… can be tested as well.
Save, import and export questions
- A question is saved in your browser by clicking on the +
- If you have already a list of questions, you can import them in the project (one question per line in a a text file)
Compare search results and answers
- A special view is provided to compare the results of different search types, vectorization models and Large Language Models. The goal is to find the best possible combination of search types, vectorization and LLMs to get the highest quality of answers.