lancedb base implementation

metagpt/document_store/lancedb_store.py

@@ -0,0 +1,87 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+@Time    : 2023/8/9 15:42
+@Author  : unkn-wn (Leon Yee)
+@File    : lancedb_store.py
+"""
+import lancedb
+import pyarrow as pa
+import pandas as pd
+
+
+class LanceStore:
+    def __init__(self, name):
+        db = lancedb.connect('./data/lancedb')
+        self.db = db
+        self.name = name
+        self.table = None
+
+    def create_table(self, columns: list):
+        # Create table given the columns as a list.
+        df = pd.DataFrame(columns=columns)
+        schema = pa.Schema.from_pandas(df)
+        schema = schema.remove_metadata()
+        schema = schema.remove(len(schema) - 1)
+
+        self.table = self.db.create_table(self.name, schema)
+
+    def search(self, query, n_results=2, metric="L2", nprobes=20, **kwargs):
+        # This assumes query is a vector embedding
+        # kwargs can be used for optional filtering
+        # .select - only searches the specified columns
+        # .where - SQL syntax filtering for metadata (e.g. where("price > 100"))
+        # .metric - specifies the distance metric to use
+        # .nprobes - values will yield better recall (more likely to find vectors if they exist) at the expense of latency.
+        results = self.table \
+            .search(query) \
+            .limit(n_results) \
+            .select(kwargs.get('select')) \
+            .where(kwargs.get('where')) \
+            .metric(metric) \
+            .nprobes(nprobes) \
+            .to_df()
+        return results
+
+    def persist(self):
+        raise NotImplementedError
+
+    def write(self, data, metadatas, ids):
+        # This function is similar to add(), but it's for more generalized updates
+        # "data" is the list of embeddings
+        # Inserts into table by expanding metadatas into a dataframe: [{'vector', 'id', 'meta', 'meta2'}, ...]
+        if self.table == None: raise Exception("Table not created yet, please use create_table([columns]) first")
+
+        documents = []
+        for i in range(len(data)):
+            row = {
+                'vector': data[i],
+                'id': ids[i]
+            }
+            row.update(metadatas[i])
+            documents.append(row)
+
+        return self.table.add(documents)
+
+    def add(self, data, metadata, _id):
+        # This function is for adding individual documents
+        # It assumes you're passing in a single vector embedding, metadata, and id
+        if self.table == None: raise Exception("Table not created yet, please use create_table([columns]) first")
+
+        row = {
+            'vector': data,
+            'id': _id
+        }
+        row.update(metadata)
+
+        return self.table.add([row])
+
+    def delete(self, _id):
+        # This function deletes a row by id.
+        # LanceDB delete syntax uses SQL syntax, so you can use "in" or "="
+        if self.table == None: raise Exception("Table not created yet, please use create_table([columns]) first")
+
+        if isinstance(_id, str):
+            return self.table.delete(f"id = '{_id}'")
+        else:
+            return self.table.delete(f"id = {_id}")

requirements.txt

@@ -10,6 +10,7 @@ faiss_cpu==1.7.4
 fire==0.4.0
 # godot==0.1.1
 # google_api_python_client==2.93.0
+lancedb==0.1.16
 langchain==0.0.231
 loguru==0.6.0
 meilisearch==0.21.0