LeomordKaly commited on
Commit
4efdd9b
·
verified ·
1 Parent(s): ccaed7e

deploy: rename Dockerfile.hf -> Dockerfile on the Space side

Browse files
Files changed (1) hide show
  1. Dockerfile +8 -4
Dockerfile CHANGED
@@ -104,10 +104,14 @@ ENV SAR_LOCAL_EMBEDDING_MODEL=BAAI/bge-m3
104
  ENV SAR_EMBEDDING_MODEL=bge-m3
105
  ENV SAR_EMBEDDING_DIM=1024
106
 
107
- # Cross-encoder reranker -- balances quality with build size. Swap to
108
- # fine_tuned + SAR_FINETUNED_RERANKER_PATH after phase 3.2 ships the
109
- # 2.3 GB checkpoint to LeomordKaly/secureagentrag-reranker-v1.
110
- ENV SAR_RERANKER_TYPE=cross_encoder
 
 
 
 
111
  ENV SAR_RERANKER_CHECKPOINT=BAAI/bge-reranker-v2-m3
112
 
113
  # Sparse retrieval -- BM25 keeps the cold path zero-dep; SPLADE adds an
 
104
  ENV SAR_EMBEDDING_MODEL=bge-m3
105
  ENV SAR_EMBEDDING_DIM=1024
106
 
107
+ # Reranker disabled for the BYOK demo. The cross-encoder adds a ~600 MB
108
+ # model + 4-5 s cold-load latency, and on a 10-doc corpus + 1-5 visitor
109
+ # uploads its top-5 cut routinely drops the visitor's own chunk before
110
+ # it reaches the grader. Pure dense+sparse RRF order is good enough on
111
+ # the demo corpus; bench shows the reranker only helps materially past
112
+ # 200+ documents per query. Re-enable for production deploys with
113
+ # SAR_RERANKER_TYPE=cross_encoder or fine_tuned.
114
+ ENV SAR_RERANKER_TYPE=none
115
  ENV SAR_RERANKER_CHECKPOINT=BAAI/bge-reranker-v2-m3
116
 
117
  # Sparse retrieval -- BM25 keeps the cold path zero-dep; SPLADE adds an