Saige is a large language model (LLM) tool that allows organisations with large sets of technical internal documentation to query these using natural language. It was born from industry feedback that conventional pipelines didn’t work well when confronted with complex documents where a lot of structured knowledge is embedded in tables, figures and equations.
Saige is able to efficiently extract and understand structured data from PDF and word documents, even in complex scenarios involving irregularly formatted tables that extend over several pages. Scientific, engineering and technology oriented businesses must also be able to rapidly verify answers so Saige embeds citation links to enable instant ‘fact checking’ from original source documents.
We can tune all elements of Saige to optimise performance for different use cases. This includes swapping different models in and out, tweaking LLM prompts, and setting appropriate trade-offs between query costs, level of detail, and processing time.
Depending on requirements, it’s possible to host Saige in the cloud, or on dedicated local infrastructure.
While Saige is built to work well with documentation, it’s possible to ingest and query records from numerous data sources – customer call records, financial transactions, effectively anything stored within a database or data lake.
For the user, the process to get Saige up and running is simple:
We start with the document sets you want to query. These can be reports you’ve created internally, reference material, or a mixture of both. Whatever folder structure you have will be mirrored by Saige, so when doing queries you can filter to narrow the scope of source material.
Document ingest may take a couple of days depending on how many documents are being ingested – 10,000 plus is achievable. During the ingest process Saige will convert documentation into HTML format and augment data by adding document and section summaries as well as structured data descriptions to improve searchability and response quality.
By default you can ask a question and your entire corpus of documents will be searched to synthesise an answer. If you want to focus the reference material to be used you can filter document sets.
The question can be a simple request for a particular discreet piece of information, or it may be quite involved and include multiple steps. If your question is complex, Saige will automatically break it down into components and build a response from multiple queries to the corpus. We use query augmentation processes and iterative steps in the background to optimise context retrieval.
A hybrid search scheme is used with vector search combined with a knowledge graph – this enables us to pinpoint the most relevant source documents and provide better context to answer a question, yielding better overall results.
You can interact with Saige just like ChatGPT or other online tools and ask it to expand on an answer or elaborate on specific aspects. Saige will provide hyperlinks with its response that will navigate you to source documentation it used to construct the answer. This is essential to check responses for accuracy, or if you need to include citations within the document you’re working on.
Response ratings help us to improve Saige; in the background Saige will continue to self improve based on the queries it receives by extracting recurring concepts and entities, optimising knowledge retrieval for these.
As Saige evolves we’re working on a Microsoft Word add-in so that the workflow is tightly integrated with the environment many clients use to create documentation.
Our medium term goal is to enable automated report completion for clients that frequently re-use report templates. In this case Saige will use similar past reports combined with the source data for the current report to ‘auto-complete’ report sections ready for editing, saving time and reducing the pain of routine report production.
If you’re interested in trying Saige within your organisation please reach out. We’re happy to discuss an evaluation beta so you can get a feel for the benefits in your unique environment.