August 9: Support direct text, Markdown and HTML ingestion; new Specification LLM strategy
Added ingestText mutation which supports direct Content ingestion of plain text, Markdown and HTML. Now, if you have pre-scraped HTML or Markdown text, you can ingest it into Graphlit without reading from a URL.
Added Specificationstrategy property, which allows customization of the LLM context when prompting a conversation. ConversationStrategy now provides Windowed and Summarized message histories, as well as configuration of the weight between existing conversation messages and Content text pages (or audio transcript segments) in the LLM context.
Added auto-summarization of extracted text and audio transcripts. There is a new Contentsummary property where a list of summary bullet points can be found. These summaries can be optionally included in the Conversation prompt context for more accurate LLM responses.
Added AzureOpenAIModels and OpenAIModels types to Specification model properties to make it easier to specify the desired LLM.
Renamed ConversationMessagedate property to timestamp
Refined the internal LLM prompts for providing content as part of Conversation context. This provides for much clearer and accurate results from the LLM.