# Overview To enhance the quality , we would be working on a customised prompt engine, further on CodeGen V2 This would include\ 1\. Token expansion , tokenization , pre processing and post processing of prompts\ 2\. Refinement with natural languages and with human feedback collection :https://arxiv.org/pdf/2303.16749.pdf\\ \ There are other techniques we are refining and plan to have this epic populated by 9th of June We would also look into having an ensemble that could further help with scalability on routing some of the traffic to Google Models , if there are areas it can perform better. Based on the preliminary evaluation , unless we do pre-process and post -processing Google Models are not superiors in reference to code completion. # Ensemble Logic ( TBD) 1. Prompt request sent to prompt engine 2. Call Gitlab Native Model 3. If results not return by X seconds call Code-chat Bison. # Additional Notes: 1. What we are trying to solve with third party is scalability ( to help with load) and incase the model does not get back anything. 2.Both Code-gen V2 and Google Model needs a lot of work for Code Completions and we are trying to build a customised prompt engine. 3.We already have the API wrapper for Google Model and there is no extra work needed , due to speed of delivery by 15th July , we would like to keep it that way. 4.We are targeting 15th July for GA complete that gives enough time for testing , changes , Infra readiness.