highperfocused/danswer
1
0
Fork 0
mirror of https://github.com/danswer-ai/danswer.git synced 2025-06-24 23:11:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

Make Boosting give more consistent scores (#406)

Browse Source
This commit is contained in:
Yuhong Sun 2023-09-06 15:50:53 -07:00 committed by GitHub
parent 2d077a9544
commit 7972c8a71e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 9 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
backend/danswer/search/semantic_search.py
View File

@ -55,19 +55,26 @@ def semantic_reranking(
query: str, query: str,
chunks: list[InferenceChunk], chunks: list[InferenceChunk],
) -> list[InferenceChunk]: ) -> list[InferenceChunk]:
model_max = 12 # These are just based on observations from model selection
model_min = -12
cross_encoders = get_default_reranking_model_ensemble() cross_encoders = get_default_reranking_model_ensemble()
sim_scores = [ sim_scores = [
encoder.predict([(query, chunk.content) for chunk in chunks]) # type: ignore encoder.predict([(query, chunk.content) for chunk in chunks]) # type: ignore
for encoder in cross_encoders for encoder in cross_encoders
] ]
cross_models_min = numpy.min(sim_scores)
shifted_sim_scores = sum( shifted_sim_scores = sum(
[enc_n_scores - numpy.min(enc_n_scores) for enc_n_scores in sim_scores] [enc_n_scores - cross_models_min for enc_n_scores in sim_scores]
) / len(sim_scores) ) / len(sim_scores)
boosts = [translate_boost_count_to_multiplier(chunk.boost) for chunk in chunks] boosts = [translate_boost_count_to_multiplier(chunk.boost) for chunk in chunks]
boosted_sim_scores = shifted_sim_scores * boosts boosted_sim_scores = shifted_sim_scores * boosts
scored_results = list(zip(boosted_sim_scores, chunks)) normalized_b_s_scores = (boosted_sim_scores + cross_models_min - model_min) / (
model_max - model_min
)
scored_results = list(zip(normalized_b_s_scores, chunks))
scored_results.sort(key=lambda x: x[0], reverse=True) scored_results.sort(key=lambda x: x[0], reverse=True)
ranked_sim_scores, ranked_chunks = zip(*scored_results) ranked_sim_scores, ranked_chunks = zip(*scored_results)