From c5babf61d72c4e3f681fda85761ef98d8f789297 Mon Sep 17 00:00:00 2001
From: yangdx <gzdaniel@me.com>
Date: Tue, 5 Aug 2025 11:38:40 +0800
Subject: [PATCH] Feat: Change embedding formats from float to base64 for
 efficiency

- Add base64 support for Jina embeddings
- Add base64 support for OpenAI embeddings
- Update env.example with new embedding options
---
 env.example            | 17 ++++++++++++-----
 lightrag/llm/jina.py   |  9 ++++++++-
 lightrag/llm/openai.py | 10 ++++++++--
 3 files changed, 28 insertions(+), 8 deletions(-)

diff --git a/env.example b/env.example
index 3967e00e..fae8122a 100644
--- a/env.example
+++ b/env.example
@@ -121,6 +121,16 @@ LLM_MODEL=gpt-4o
 LLM_BINDING_HOST=https://api.openai.com/v1
 LLM_BINDING_API_KEY=your_api_key
 
+### Optional for Azure
+# AZURE_OPENAI_API_VERSION=2024-08-01-preview
+# AZURE_OPENAI_DEPLOYMENT=gpt-4o
+
+### Openrouter example
+# LLM_MODEL=google/gemini-2.5-flash
+# LLM_BINDING_HOST=https://openrouter.ai/api/v1
+# LLM_BINDING_API_KEY=your_api_key
+# LLM_BINDING=openai
+
 ### Most Commont Parameters for Ollama Server
 ### Time out in seconds, None for infinite timeout
 TIMEOUT=240
@@ -132,14 +142,11 @@ OLLAMA_LLM_NUM_CTX=32768
 # OLLAMA_LLM_TEMPERATURE=0.85
 ### see also env.ollama-binding-options.example for fine tuning ollama
 
-### Optional for Azure
-# AZURE_OPENAI_API_VERSION=2024-08-01-preview
-# AZURE_OPENAI_DEPLOYMENT=gpt-4o
 
 ####################################################################################
 ### Embedding Configuration (Should not be changed after the first file processed)
 ####################################################################################
-### Embedding Binding type: openai, ollama, lollms, azure_openai, jina
+### Embedding Binding type: ollama, openai, azure_openai, jina,  lollms
 
 ### see also env.ollama-binding-options.example for fine tuning ollama
 EMBEDDING_BINDING=ollama
@@ -149,7 +156,7 @@ EMBEDDING_BINDING_API_KEY=your_api_key
 # If the embedding service is deployed within the same Docker stack, use host.docker.internal instead of localhost
 EMBEDDING_BINDING_HOST=http://localhost:11434
 
-### OpenAI compatible
+### OpenAI compatible (VoyageAI embedding openai compatible)
 # EMBEDDING_BINDING=openai
 # EMBEDDING_MODEL=text-embedding-3-large
 # EMBEDDING_DIM=3072
diff --git a/lightrag/llm/jina.py b/lightrag/llm/jina.py
index 5a1b59fb..af26f1bd 100644
--- a/lightrag/llm/jina.py
+++ b/lightrag/llm/jina.py
@@ -8,6 +8,7 @@ if not pm.is_installed("tenacity"):
     pm.install("tenacity")
 
 import numpy as np
+import base64
 import aiohttp
 from tenacity import (
     retry,
@@ -82,6 +83,7 @@ async def jina_embed(
         "model": "jina-embeddings-v4",
         "task": "text-matching",
         "dimensions": dimensions,
+        "embedding_type": "base64",
         "input": texts,
     }
 
@@ -108,7 +110,12 @@ async def jina_embed(
                 f"Jina API returned {len(data_list)} embeddings for {len(texts)} texts"
             )
 
-        embeddings = np.array([dp["embedding"] for dp in data_list])
+        embeddings = np.array(
+            [
+                np.frombuffer(base64.b64decode(dp["embedding"]), dtype=np.float32)
+                for dp in data_list
+            ]
+        )
         logger.debug(f"Jina embeddings generated: shape {embeddings.shape}")
 
         return embeddings
diff --git a/lightrag/llm/openai.py b/lightrag/llm/openai.py
index 981669b6..5a52e767 100644
--- a/lightrag/llm/openai.py
+++ b/lightrag/llm/openai.py
@@ -34,6 +34,7 @@ from lightrag.types import GPTKeywordExtractionFormat
 from lightrag.api import __api_version__
 
 import numpy as np
+import base64
 from typing import Any, Union
 
 from dotenv import load_dotenv
@@ -472,6 +473,11 @@ async def openai_embed(
 
     async with openai_async_client:
         response = await openai_async_client.embeddings.create(
-            model=model, input=texts, encoding_format="float"
+            model=model, input=texts, encoding_format="base64"
+        )
+        return np.array(
+            [
+                np.frombuffer(base64.b64decode(dp.embedding), dtype=np.float32)
+                for dp in response.data
+            ]
         )
-        return np.array([dp.embedding for dp in response.data])