August 8: Support for LLM-based document extraction, .NET SDK, bug fixes
New Features
💡 Graphlit now supports LLM-based document preparation, using vision-capable models such as OpenAI GPT-4o and Anthropic Sonnet 3.5. This is available via the
MODEL_DOCUMENT
preparation service type, and you can assign a customspecification
object and bring your own LLM keys.Added
identifier
property toContent
object for mapping content to external database identifiers. This is supported for content filtering as well.Added support for Claude 3 vision models for image-based entity extraction, using the
MODEL_IMAGE
entity extraction service.Added context augmentation to conversations, via the
augmentedFilter
property on theConversation
object. Any content which matches this augmented filter will be injected into the LLM prompt content, without needing to be related by vector similarity to the user prompt. This is useful for specifying domain knowledge which should always be referenced by the RAG pipeline.Added support for the latest snapshot of OpenAI GPT-4o, with the model enum
GPT4O_128K_20240806.
Added reranking of related entities, when preparing the LLM prompt context for GraphRAG. If reranking is enabled, the metadata from the related entities will be reranked with the same reranker assigned to the conversation specification.
⚡ We have changed the type of the
duration
field in theAudioMetadata
andVideoMetadata
types to beTimeSpan
rather thanstring
, as to be more consistent with the rest of the API data model.
Bugs Fixed
GPLA-2884: Support retry on HTTP 529 (Overloaded) error from Anthropic API.
Last updated