flashrag-dev 0.3.0.dev20260410__tar.gz → 0.3.0.dev20260809__tar.gz

{flashrag_dev-0.3.0.dev20260410/flashrag_dev.egg-info → flashrag_dev-0.3.0.dev20260809}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flashrag_dev
-Version: 0.3.0.dev20260410
+Version: 0.3.0.dev20260809
 Summary: A library for efficient Retrieval-Augmented Generation research
 Home-page: https://github.com/RUC-NLPIR/FlashRAG
 Author: Jiajie Jin, Yutao Zhu, Chenghao Zhang, Xinyu Yang, Zhicheng Dou

{flashrag_dev-0.3.0.dev20260410 → flashrag_dev-0.3.0.dev20260809}/flashrag/pipeline/active_pipeline.py

@@ -855,7 +855,7 @@ class SelfAskPipeline(BasicPipeline):
                 self.P_INS
                 + "\n"
                 + self.format_reference(retrieval_result)
-                + f"\nQuesiton: {question}"
+                + f"\nQuestion: {question}"
                 + "\nAre follow up questions needed here: "
                 + follow_ups
                 + "\n"
@@ -891,7 +891,7 @@ class SelfAskPipeline(BasicPipeline):
             if "So the final answer is: " in gen_out:
                 res = (
                     self.format_reference(retrieval_result)
-                    + f"\nQuesiton: {question}"
+                    + f"\nQuestion: {question}"
                     + "\nAre follow up questions needed here: "
                     + follow_ups
                     + "\n"
@@ -904,7 +904,7 @@ class SelfAskPipeline(BasicPipeline):
         if not early_exit:
             res = (
                 self.format_reference(retrieval_result)
-                + f"\nQuesiton: {question}"
+                + f"\nQuestion: {question}"
                 + "\nAre follow up questions needed here: "
                 + follow_ups
                 + "\n"

{flashrag_dev-0.3.0.dev20260410 → flashrag_dev-0.3.0.dev20260809}/flashrag/pipeline/pipeline.py

@@ -141,9 +141,10 @@ class ConditionalPipeline(BasicPipeline):
 
         self.judger = get_judger(config)
         if generator is None:
-            self.generator = get_generator(config)
+            generator = get_generator(config)
         if retriever is None:
-            self.retriever = get_retriever(config)
+            retriever = get_retriever(config)
+
         self.generator = generator
         self.retriever = retriever
 

{flashrag_dev-0.3.0.dev20260410 → flashrag_dev-0.3.0.dev20260809}/flashrag/retriever/index_builder.py

@@ -481,15 +481,24 @@ class Index_Builder:
         else:
             memmap[:] = all_embeddings
 
-    def encode_all(self):
+    def encode_all(self, hidden_size=None):
         encode_data = [item["contents"] for item in self.corpus]
         if self.gpu_num > 1:
             print("Use multi gpu!")
             self.batch_size = self.batch_size * self.gpu_num
-            all_embeddings = self.encoder.multi_gpu_encode(encode_data, batch_size=self.batch_size, is_query=False)
-        else:
-            all_embeddings = self.encoder.encode(encode_data, batch_size=self.batch_size, is_query=False)
-
+            return self.encoder.multi_gpu_encode(encode_data, batch_size=self.batch_size, is_query=False)
+        if self.use_sentence_transformer:
+            return self.encoder.encode(encode_data, batch_size=self.batch_size, is_query=False)
+
+        # Stream each batch straight into the on-disk memmap instead of buffering the whole corpus in RAM.
+        corpus_size = len(encode_data)
+        all_embeddings = np.memmap(
+            self.embedding_save_path, shape=(corpus_size, hidden_size), mode="w+", dtype=np.float32
+        )
+        for i in tqdm(range(0, corpus_size, self.batch_size), desc="Encoding process: ", disable=self.encoder.silent):
+            batch_emb = self.encoder.single_batch_encode(encode_data[i : i + self.batch_size], is_query=False)
+            all_embeddings[i : i + batch_emb.shape[0]] = batch_emb
+        all_embeddings.flush()
         return all_embeddings
 
     def encode_all_clip(self):
@@ -549,12 +558,14 @@ class Index_Builder:
             )
             hidden_size = self.encoder.model.config.hidden_size
 
+        streamed_embedding_to_disk = False
         if self.embedding_path is not None:
             corpus_size = len(self.corpus)
             all_embeddings = self._load_embedding(self.embedding_path, corpus_size, hidden_size)
         else:
-            all_embeddings = self.encode_all_clip() if self.is_clip else self.encode_all()
-            if self.save_embedding:
+            streamed_embedding_to_disk = not self.is_clip and self.gpu_num <= 1 and not self.use_sentence_transformer
+            all_embeddings = self.encode_all_clip() if self.is_clip else self.encode_all(hidden_size)
+            if self.save_embedding and not streamed_embedding_to_disk:
                 self._save_embedding(all_embeddings)
             del self.corpus
 
@@ -583,6 +594,10 @@ class Index_Builder:
             if os.path.exists(self.index_save_path):
                 print("The index file already exists and will be overwritten.")
             self.save_faiss_index(all_embeddings, self.faiss_type, self.index_save_path)
+
+        if streamed_embedding_to_disk and not self.save_embedding:
+            del all_embeddings
+            os.remove(self.embedding_save_path)
         print("Finish!")
 
     def save_faiss_index(

flashrag_dev-0.3.0.dev20260809/flashrag/version.py

@@ -0,0 +1 @@
+__version__ = "0.3.0dev20260809"

{flashrag_dev-0.3.0.dev20260410 → flashrag_dev-0.3.0.dev20260809/flashrag_dev.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flashrag_dev
-Version: 0.3.0.dev20260410
+Version: 0.3.0.dev20260809
 Summary: A library for efficient Retrieval-Augmented Generation research
 Home-page: https://github.com/RUC-NLPIR/FlashRAG
 Author: Jiajie Jin, Yutao Zhu, Chenghao Zhang, Xinyu Yang, Zhicheng Dou

flashrag_dev-0.3.0.dev20260410/flashrag/version.py

@@ -1 +0,0 @@
-__version__ = "0.3.0dev20260410"