Fixed openai embeddings to be safe by batching them based on token si…

…ze calculation. (langchain-ai#991)

I modified the logic of the batch calculation for embedding according to
this cookbook

https://github.com/openai/openai-cookbook/blob/main/examples/Embedding_long_inputs.ipynb

langchain/embeddings/openai.py

@@ -1,6 +1,7 @@
 """Wrapper around OpenAI embedding models."""
 from typing import Any, Dict, List, Optional
 
+import numpy as np
 from pydantic import BaseModel, Extra, root_validator
 
 from langchain.embeddings.base import Embeddings
@@ -24,6 +25,7 @@ class OpenAIEmbeddings(BaseModel, Embeddings):
     client: Any  #: :meta private:
     document_model_name: str = "text-embedding-ada-002"
     query_model_name: str = "text-embedding-ada-002"
+    embedding_ctx_length: int = -1
     openai_api_key: Optional[str] = None
 
     class Config:
@@ -69,11 +71,62 @@ def validate_environment(cls, values: Dict) -> Dict:
             )
         return values
 
+    # please refer to
+    # https://github.com/openai/openai-cookbook/blob/main/examples/Embedding_long_inputs.ipynb
+    def _get_len_safe_embeddings(
+        self, texts: List[str], *, engine: str, chunk_size: int = 1000
+    ) -> List[List[float]]:
+        embeddings: List[List[float]] = [[] for i in range(len(texts))]
+        try:
+            import tiktoken
+
+            tokens = []
+            indices = []
+            encoding = tiktoken.model.encoding_for_model(self.document_model_name)
+            for i, text in enumerate(texts):
+                # replace newlines, which can negatively affect performance.
+                text = text.replace("\n", " ")
+                token = encoding.encode(text)
+                for j in range(0, len(token), self.embedding_ctx_length):
+                    tokens += [token[j : j + self.embedding_ctx_length]]
+                    indices += [i]
+
+            batched_embeddings = []
+            for i in range(0, len(tokens), chunk_size):
+                response = self.client.create(
+                    input=tokens[i : i + chunk_size], engine=self.document_model_name
+                )
+                batched_embeddings += [r["embedding"] for r in response["data"]]
+
+            results: List[List[List[float]]] = [[] for i in range(len(texts))]
+            lens: List[List[int]] = [[] for i in range(len(texts))]
+            for i in range(len(indices)):
+                results[indices[i]].append(batched_embeddings[i])
+                lens[indices[i]].append(len(batched_embeddings[i]))
+
+            for i in range(len(texts)):
+                average = np.average(results[i], axis=0, weights=lens[i])
+                embeddings[i] = (average / np.linalg.norm(average)).tolist()
+
+            return embeddings
+
+        except ImportError:
+            raise ValueError(
+                "Could not import tiktoken python package. "
+                "This is needed in order to for OpenAIEmbeddings. "
+                "Please it install it with `pip install tiktoken`."
+            )
+
     def _embedding_func(self, text: str, *, engine: str) -> List[float]:
         """Call out to OpenAI's embedding endpoint."""
         # replace newlines, which can negatively affect performance.
-        text = text.replace("\n", " ")
-        return self.client.create(input=[text], engine=engine)["data"][0]["embedding"]
+        if self.embedding_ctx_length > 0:
+            return self._get_len_safe_embeddings([text], engine=engine)[0]
+        else:
+            text = text.replace("\n", " ")
+            return self.client.create(input=[text], engine=engine)["data"][0][
+                "embedding"
+            ]
 
     def embed_documents(
         self, texts: List[str], chunk_size: int = 1000
@@ -89,13 +142,16 @@ def embed_documents(
             List of embeddings, one for each text.
         """
         # handle large batches of texts
-        results = []
-        for i in range(0, len(texts), chunk_size):
-            response = self.client.create(
-                input=texts[i : i + chunk_size], engine=self.document_model_name
+        if self.embedding_ctx_length > 0:
+            return self._get_len_safe_embeddings(
+                texts, engine=self.document_model_name, chunk_size=chunk_size
             )
-            results += [r["embedding"] for r in response["data"]]
-        return results
+        else:
+            responses = [
+                self._embedding_func(text, engine=self.document_model_name)
+                for text in texts
+            ]
+            return responses
 
     def embed_query(self, text: str) -> List[float]:
         """Call out to OpenAI's embedding endpoint for embedding query text.

tests/integration_tests/embeddings/test_openai.py

@@ -15,6 +15,7 @@ def test_openai_embedding_documents_multiple() -> None:
     """Test openai embeddings."""
     documents = ["foo bar", "bar foo", "foo"]
     embedding = OpenAIEmbeddings()
+    embedding.embedding_ctx_length = 8191
     output = embedding.embed_documents(documents, chunk_size=2)
     assert len(output) == 3
     assert len(output[0]) == 1536