fix: Cursor Agent Review found DevisionByZero in search.py if no data collection selected, and Claude also put the same guard in UDR

chore: refactoring inline imports

fix: the same bug in TTD-DR what UDR had: passing the comma-separated collection string directly to Milvus instead of iterating over individual collections #26

fix: refactor references resolving logic duplicated code + finding one more place where it was nto applied #26 #1 #6 #7 #8

fix: references are co-mingled when multiple data collections are selected #26 #1 #6 #7 #8

chore: cleaned up 5 Milvus Pulsar nodes started 34d ago, the newer Milvus uses woodpecker as the messaging #25

fix: a somke test revealed that moving the nvext guided_json into the root level of the payload and removing the extra_body wrapper is a complete f-up! #21 #25

fix: TTD_DR errors out because of confidence scores meant to be 1-100, but LLM produces float [0.0-1.0].

feat: documenting how we got TensorRT LLM serving working with g5.8xlarge instance: 64GB system RAM prevented OOM during build #25

chore: add tests for the NVIDIA guided_json calls #21 #25

chore: avoid inline imports in the guided_json refactor files #21 #25

fix: refactor structured input (NVIDIA guided_json calls) after all the updates #21 #25

fix: Explore Nemotron Nano 9B V2 vs Nano 8B V1, also BF8 and BF16, KV cache vs 16KB context size #25

feat: Upgrading Karpenter from v0.32.0 to v1.0.x for AL2023 support #24, w Helm chart #24

fix: Upgrade from Kubernetes 1.31 (#25) to latest 1.34, which also require upgrade from AL2 to AL2023, fixes #24