November 10: Support for web search, multi-turn content revision, Deepgram transcription language
Last updated
Last updated
Graphlit now supports web search with the searchWeb mutation. You can select the search service, either Tavily or Exa.AI, and provide the search query and number of search results to be returned. This is different than the web search feed, in that searchWeb returns the relevant text from the web page and the web page URL from each search hit, but does not ingest each of the web pages. This new mutation is optimized to be used from within an LLM tool.
Graphlit now supports multi-turn revision of content with the reviseContent mutation. You can provide an LLM prompt and a content reference, along with an optional specification. This can be used for summarizing any content (documents, web pages, audio transcripts, etc.), and having a multi-turn conversation with the LLM to revise the output from the LLM. Internally, this creates a conversation locked to a single piece of content. This works especially well with the OpenAI o1-preview and o1-mini models, because they provide a longer LLM output from each turn.
Graphlit now supports the configuration of the Deepgram transcription language, and whether detectLanguage is enabled in DeepgramAudioPreparationPropertiesInput. Language detection is now enabled by default, and can be disabled by setting detectLanguage to false.
We have added a requireTool option to promptConversation mutation, so you can control whether the LLM must call one of the provided tool, or if tool calling is optional.
For accounts created after Nov 8, 2024, we have lowered the credits quota on the Free tier from 1000 credits to 100 credits, and now offer unlimited feeds on the Hobby Tier.
The Graphlit Data API will now return HTTP 402 (Payment Required) when you have exceeded the credits quota on the free tier. You must upgrade to the Hobby Tier (or higher) to continue using the API, once the credits quota has been reached.