Update README.md

README.md

@@ -92,8 +92,77 @@ It created a nice interactive map with tooltips.
 
 ![image/png](https://cdn-uploads.huggingface.co/production/uploads/64c4da8719565937fb268b32/QIolrK2nlrENnkWlE6TFh.png)
 
-## Remote Semantic Search 
+## Semantic Search 
 
+### Local and Remote Semantic Search with LanceDB
+
+#### Reading a LanceDB index
+
+You can use the text LanceDB index I created for Italy in this way: 
+
+Download the index locally to get the fastest results (3.4GB).
+
+Load the model and connect to the local DB:
+
+```python
+from model2vec import StaticModel
+import numpy as np
+import lancedb
+
+model = StaticModel.from_pretrained("minishlab/m2v_multilingual_output")
+
+db = lancedb.connect("italy_lancedb")
+table = db.open_table("foursquare")
+```
+
+Create the query vector: 
+
+```python
+query = "ski and snowboard"
+query_vector = model.encode(query),
+query_vector = np.array(query_vector).astype(np.float32)[0]
+```
+
+Fire the query. On my M3 Max I usually get query times of ~5ms 
+
+```python
+table.search(query_vector).limit(10).select(["poi_name", "latitude", "longitude"]).to_pandas().sort_values("_distance",ascending=False)
+```
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/64c4da8719565937fb268b32/tS7cbe_75H0CM0imx23Km.png)
+
+You can host the index on s3 or similar, but with a massive increase in query latency (in the hundreds of ms, but under 1s). See LanceDB docs for more info.
+
+#### Writing a LanceDB index
+
+```python
+import geopandas as gpd
+
+data = gpd.read_parquet("foursquare_places_italy_embeddings.parquet")
+
+import lancedb
+import geopandas as gpd
+import pyarrow as pa
+
+# Connect to LanceDB
+db = lancedb.connect("italy_lancedb")
+
+# Extract lat/lon from geometry
+data['latitude'] = data['geometry'].y
+data['longitude'] = data['geometry'].x
+
+data = data.rename(columns={"name": "poi_name", "embeddings": "vector"})  # standardize vector name to "vector", rename column to avoid conflicts
+data = data[['poi_name', 'vector', 'latitude', 'longitude']] # select valid columns
+
+# Create the table
+table = db.create_table("foursquare", data=data)
+
+# Create a Vector Index (Optional but HIGHLY recommended)
+table.create_index(vector_column_name="vector", metric="cosine", num_sub_vectors=16) # do not use accelerator="mps" on Mac for tens of millions of rows, only for much larger DBs
+```
+
+
+### Remote Semantic Search with DuckDB
 You can even perform semantic search remotely without downloading the whole file.
 Without using any index on the data (like HNSW or ANN etc.) but by simply brute forcing, it takes around 3 mins on my machine to query the example file for Italy remotely. 
 It weighs ~5Gb and consists of 3.029.191 rows. I used https://huggingface.co/minishlab/M2V_multilingual_output as multilingual embeddings.