Skip to content

100+ AI Agents, Agent Skills and RAG Apps - Free and Open Source.

Fix Beifong FAISS search to convert L2 distance to cosine similarity

midasstack53b1ab562022
diff --git a/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tests/embedding_search_test.py b/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tests/embedding_search_test.py
index cbe51a99..b3828e74 100644
--- a/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tests/embedding_search_test.py
+++ b/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tests/embedding_search_test.py
@@ -12,6 +12,12 @@ EMBEDDING_MODEL = "text-embedding-3-small"
FAISS_INDEX_PATH, FAIS_MAPPING_PATH = get_faiss_db_path()
+def l2_distance_to_cosine_similarity(distance: float) -> float:
+ """Convert FAISS L2 distance to cosine similarity for unit-normalized embeddings."""
+ similarity = 1.0 - (float(distance) ** 2) / 2.0
+ return float(max(0.0, min(1.0, similarity)))
+
+
def generate_query_embedding(client, query_text, model=EMBEDDING_MODEL):
try:
response = client.embeddings.create(input=query_text, model=model)
@@ -93,7 +99,7 @@ def search_articles(
results = get_article_details(tracking_db_path, result_article_ids)
for i, result in enumerate(results):
distance = float(distances[0][i])
- similarity = float(np.exp(-distance))
+ similarity = l2_distance_to_cosine_similarity(distance)
result["distance"] = distance
result["similarity"] = similarity
result["score"] = similarity
diff --git a/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tools/embedding_search.py b/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tools/embedding_search.py
index d7333ef6..fbbdcf3b 100644
--- a/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tools/embedding_search.py
+++ b/advanced_ai_agents/multi_agent_apps/ai_news_and_podcast_agents/beifong/tools/embedding_search.py
@@ -12,6 +12,17 @@ import json
EMBEDDING_MODEL = "text-embedding-3-small"
+def l2_distance_to_cosine_similarity(distance: float) -> float:
+ """Convert FAISS L2 distance to cosine similarity for unit-normalized embeddings.
+
+ OpenAI text-embedding-3 vectors are length-1, so:
+ ||a - b||^2 = 2 - 2 * cos(a, b)
+ cos(a, b) = 1 - ||a - b||^2 / 2
+ """
+ similarity = 1.0 - (float(distance) ** 2) / 2.0
+ return float(max(0.0, min(1.0, similarity)))
+
+
def generate_query_embedding(query_text, model=EMBEDDING_MODEL):
try:
api_key = load_api_key("OPENAI_API_KEY")
@@ -116,7 +127,7 @@ def embedding_search(agent: Agent, prompt: str) -> str:
for i, idx in enumerate(indices[0]):
if idx >= 0 and idx < len(id_map):
distance = float(distances[0][i])
- similarity = float(np.exp(-distance)) if distance > 0 else 0
+ similarity = l2_distance_to_cosine_similarity(distance)
if similarity >= similarity_threshold:
article_id = id_map[idx]
results_with_metrics.append((idx, distance, similarity, article_id))