# September 20: Paid subscription plans; support for custom observed entities & Azure OpenAI GPT-4

### New Features

* :fire: Graphlit now supports paid [Hobby, Starter and Growth](https://www.graphlit.com/#pricing) tiers for projects, in addition to the existing Free tier.  Starting at $49/mo, plus $0.10/credit for usage, we now support higher quota based on your subscribed tier.   By providing a payment method for your organization in the Developer Portal, you can upgrade each project individually to the tier that fits your application's needs.
* :bulb: Added GraphQL mutations for the creation, update and deletion of observed entities (i.e. Person, Organization, Place, Product, Event, Label, Category).
* :bulb: Added new observed entity types to knowledge graph: **Repo** (i.e. Git repo), **Software**.
* :bulb: Added `searchType` and `numberSimilar` fields to Specification object for configuring semantic search in conversations.   In situations where the user prompt is limited in length, `HYBRID` search type can provide better semantic search results for the prompt context.
* :bulb: Added support for the Azure OpenAI GPT-4 model.
* Added support for project `quota` field.  Project quotas are based on the subscribed pricing tier.   Quota limits are now applied as content is ingested, and as feeds and conversations are created.
* Added `contentLimit` to conversation strategy object to limit the number of semantic search content results which are formatted into prompt context.
* Better relevance ranking on semantic search results when formatting prompt context in conversations.
* :information\_source: Free tier has updated quota: 1GB storage, 100 contents, 3 feeds and 10 conversations.
* :zap: Now using the [Deepgram Nova-2 audio transcription model,](https://deepgram.com/learn/nova-2-speech-to-text-api?utm_source=twitter\&utm_campaign=0923+Nova-2\&utm_medium=organic+social) which is 18% more accurate, and 5-40x faster.

### Bugs Fixed

* GPLA-1373: Failed to extract multiple text pages from DOCX without page breaks.  Now we support token-aware page chunking.
* GPLA-1377: Failed during semantic search with no content results, when prompting conversation.
* GPLA-1415: Failed when user prompt couldn't generate text embeddings.