Merge pull request #571 from valentimarco/develop

Change Fast Embedder adapter using langchain class

.gitignore

@@ -32,3 +32,6 @@ core/cat/plugins/*
 # tests plugin folder
 plugin_folder
 
+# cache embedder
+core/local_cache/*
+

core/cat/factory/custom_embedder.py

@@ -65,22 +65,3 @@ def embed_query(self, text: str) -> List[float]:
         ret.raise_for_status()
         return ret.json()['data'][0]['embedding']
 
-class CustomFastembedEmbeddings(Embeddings):
-    """Use Fastembed for embedding.
-    """
-    def __init__(self, url, model,max_length) -> None:
-        self.url = url
-        output = httpx.post(f"{url}/embeddings", json={"model": model, "max_length": max_length}, follow_redirects=True, timeout=None)
-        output.raise_for_status()
-
-
-    def embed_documents(self, texts: List[str]):
-        ret = httpx.post(f"{self.url}/embeddings/document", json={"document": texts}, timeout=None)
-        ret.raise_for_status()
-        return ret.json()
-
-    def embed_query(self, text: str) -> List[float]:
-        ret = httpx.post(f"{self.url}/embeddings/prompt", json={"prompt": text}, timeout=None)
-        ret.raise_for_status()
-        return ret.json()
-

core/cat/factory/embedder.py

@@ -1,8 +1,8 @@
 from typing import Type
 import langchain
 from pydantic import BaseModel, ConfigDict
-
-from cat.factory.custom_embedder import CustomFastembedEmbeddings, DumbEmbedder, CustomOpenAIEmbeddings
+from langchain.embeddings.fastembed import FastEmbedEmbeddings
+from cat.factory.custom_embedder import DumbEmbedder, CustomOpenAIEmbeddings
 
 
 # Base class to manage LLM configuration.
@@ -107,19 +107,20 @@ class EmbedderCohereConfig(EmbedderSettings):
     )
 
 
-class EmbedderFastEmbedConfig(EmbedderSettings):
-    url: str
-    model: str = "intfloat/multilingual-e5-large"
-    max_length: int = 512
-
-    _pyclass: Type = CustomFastembedEmbeddings
+class EmbedderQdrantFastEmbedConfig(EmbedderSettings):
+    model_name: str = "BAAI/bge-base-en"
+    max_length: int = 512 # Unknown behavior for values > 512.
+    doc_embed_type: str = "passage" # as suggest on fastembed documentation, "passage" is the best option for documents.
+    
+    _pyclass: Type = FastEmbedEmbeddings
 
     model_config = ConfigDict(
         json_schema_extra = {
-            "humanReadableName": "Fast Embedder",
-            "description": "Configuration for Fast embeddings",
+            "humanReadableName": "Qdrant FastEmbed (Local)",
+            "description": "Configuration for Qdrant FastEmbed",
         }
     )
+
 
 
 SUPPORTED_EMDEDDING_MODELS = [
@@ -129,7 +130,7 @@ class EmbedderFastEmbedConfig(EmbedderSettings):
     EmbedderOpenAIConfig,
     EmbedderAzureOpenAIConfig,
     EmbedderCohereConfig,
-    EmbedderFastEmbedConfig
+    EmbedderQdrantFastEmbedConfig
 ]
 
 

core/pyproject.toml

@@ -15,7 +15,7 @@ dependencies = [
     "pandas==1.5.3",
     "scikit-learn==1.2.1",
     "qdrant_client==1.6.6",
-    "langchain==0.0.315",
+    "langchain==0.0.336",
     "openai==0.27.5",
     "cohere==4.0.4",
     "huggingface-hub==0.13.2",
@@ -37,7 +37,8 @@ dependencies = [
     "perflint",
     "pylint-actions",
     "pytest",
-    "httpx"
+    "httpx",
+    "fastembed==0.1.1"
 ]
 
 [tool.coverage.run]