May 15: Support for GraphRAG, OpenAI GPT-4o model, performance improvements and bug fixes

New Features

• 💡 Graphlit now supports GraphRAG, where the extracted entities in the knowledge graph can be added as additional context to your RAG con,versation. Also, with GraphRAG, entities can be extracted from the user prompt, and used as additional content filters - or can be used to query related content sources, which are combined with the vector search results. This can be configured by specifying your graphStrategy in the Specification object.
• 💡 Graphlit now supports LLM revisions within RAG conversations, where the LLM can be prompted to revise its initial completion response. From our testing, this has been shown to provide 35% more output tokens with higher quality responses. This can be configured by specifying your revisionStrategy, and you can use our built-in revision prompt, or provide a custom one, and specify how many revisions you want the LLM to make.
• 💡 Graphlit now supports the new OpenAI GPT-4o model for RAG conversations.
• ⚡ We have changed the default model for Conversations to be OpenAI GPT-4o, from Azure OpenAI GPT-3.5 16k. This provides faster performance and better quality output.
• Add graph to promptConversation response, so you can visualize or leverage the nodes and edges of the knowledge graph, resulting from the content retrieval. For example, if a Person and Organization were observed in the cited content sources used by the RAG pipeline, you will get back those entities and their relationship (such as Person 'works-for' Organization).

• Expanded the enriched data from WIkipedia to include the long description of an entity.

• Added getSharePointLibraries, getSharePointFolders, and getOneDriveFolders queries to the API, which can be used to enumerate the storage services. This makes locating the SharePoint libraryId easier, for example.

• Added getTeams and getTeamsChannels queries to the API for enumerating Microsoft Teams workspaces.

• Added extractedCount to the entity extraction connector to limit the number of extracted entities, per entity type. I.e. if extracted count is 10, it will extract at most ten each of Persons, Organizations, etc.

• 🔥 We have improved performance in several areas: creation of observations after entity extraction, access to cloud storage, rendering the RAG context.
• 🔥 We have optimized the LLM entity extraction process to identify more properties, as well as entity-to-entity relationships.

Bugs Fixed

• GPLA-2652: Not extracting text from HTML in RSS post

• GPLA-2627: Limit filter only returning half the results

• GPLA-2613: Not properly extracting structured text from JSON/XML formats