December 10: Support for OpenAI GPT-4 Turbo, Llama 2 and Mistral models; query by example, bug fixes

New Features

• 💡 Graphlit now supports the OpenAI GPT-4 Turbo 128k model, both in Azure OpenAI and native OpenAI services. Added new model enum GPT4_TURBO_VISION_128K.

• 💡 Graphlit now supports Llama 2 7b, 13b, 70b models and Mistral 7b model, via Replicate. Developers can use their own Replicate API key, or be charged as credits for Graphlit usage.

• 💡 Graphlit now supports the Anthropic Claude 2.1 model. Added new model enum CLAUDE_2_1.

• 💡 Graphlit now supports the OpenAI GPT-4 Vision model for image descriptions and text extraction. Added new model enum GPT4_TURBO_VISION_128K. See usage example in "Multimodal RAG" blog post.

• Added query by example to contents query. Developers can specify one or more example contents, and query will use vector embeddings to return similar contents.

• Added query by example to conversations query. Developers can specify one or more example conversations, and query will use vector embeddings to return similar conversations.

• Added vector search support for conversations queries. Developers can provide search text which will use vector embeddings to return similar conversations.

• Added promptSpecifications mutation for directly prompting multiple models. This can be used to evaluate prompts against multiple models or compare different specification parameters in parallel.

• Added promptStrategy field to Specification, which supports multiple strategy types for preprocessing the prompt before being sent to the LLM model. For example, REWRITE prompt strategy will ask LLM to rewrite the incoming user prompt based on the previous conversation messages.

• Added suggestConversation mutation, which returns a list of suggested followup questions based on the specified conversation and related contents. This can be used to auto-suggest questions for chatbot users.

• Added new summarization types: CHAPTERS, QUESTIONS and POSTS. See usage examples in the "LLMs for Podcasters" blog post.

• Added versioned model enums such as GPT4_0613 and GPT35_TURBO_16K_1106. Without version specified, such as GPT35_TURBO_16K, Graphlit will use the latest production model version, as defined by the LLM vendor.

• Added lookupContents query to get multiple contents by id in one query.

• ⚡ In Content type, headline field was renamed to headlines and now returns an array of strings.

• ⚡ Entity names are now limited to 1024 characters. Names will be truncated if they exceed the maximum length.

• ⚡ In SummarizationTypes enum, BULLET_POINTS was renamed to BULLETS.

• ⚡ In ProjectStorage type, originalTotalSize was renamed to totalSize, and totalRenditionSize field was added. totalSize is the sum of the ingested source file sizes, and totalRenditionSize is the sum of the source file sizes and any derived rendition sizes.

• ⚡ In ConversationStrategy type, strategyType was renamed to type for consistency with rest of data model.

• ⚡ In Specification type, optimizeSearchConversation was removed, and now is handled by OPTIMIZE_SEARCH prompt strategy.

Bugs Fixed

• GPLA-1725: Should ignore RSS.xml from web feed sitemap

• GPLA-1726: GPT-3.5 Turbo 16k LLM is adding "Citation #" to response

• GPLA-1698: Workflow not applied to link-crawled content

• GPLA-1692: Mismatched project storage total size, when some content has errored

• GPLA-1237: Add relevance threshold for semantic search