flashrag-dev 0.1.3.dev20241115__tar.gz → 0.1.3.dev20241119__tar.gz

{flashrag-dev-0.1.3.dev20241115/flashrag_dev.egg-info → flashrag-dev-0.1.3.dev20241119}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: flashrag-dev
-Version: 0.1.3.dev20241115
+Version: 0.1.3.dev20241119
 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-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/dataset/dataset.py

@@ -174,9 +174,12 @@ class Dataset:
         """Save the dataset into the original format."""
 
         save_data = [item.to_dict() for item in self.data]
-
+        def custom_serializer(obj):
+            if isinstance(obj, np.float32):  
+                return float(obj)           
+            raise TypeError(f"Type {type(obj)} not serializable")
         with open(save_path, "w", encoding="utf-8") as f:
-            json.dump(save_data, f, indent=4)
+            json.dump(save_data, f, indent=4, default=custom_serializer)
 
     def __str__(self) -> str:
         """Return a string representation of the dataset with a summary of items."""

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/dataset/utils.py

@@ -13,6 +13,8 @@ def convert_numpy(obj: Union[Dict, list, np.ndarray, np.generic]) -> Any:
         return obj.tolist()  # Convert numpy arrays to lists
     elif isinstance(obj, (np.integer, np.floating)):
         return obj.item()  # Convert numpy scalars to native Python scalars
+    elif isinstance(obj, np.float32):
+        return float(obj)
     else:
         return obj  # Return the object as-is if it's neither a dict, list, nor numpy type
 

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/pipeline/active_pipeline.py

@@ -898,12 +898,12 @@ class SelfAskPipeline(BasicPipeline):
         dataset = self.evaluate(dataset, do_eval=do_eval, pred_process_fun=pred_process_fun)
         return dataset
 
-
+    
 class IRCOTPipeline(BasicPipeline):
     IRCOT_INSTRUCTION = 'You serve as an intelligent assistant, adept at facilitating users through complex, multi-hop reasoning across multiple documents. This task is illustrated through demonstrations, each consisting of a document set paired with a relevant question and its multi-hop reasoning thoughts. Your task is to generate one thought for current step, DON\'T generate the whole thoughts at once! If you reach what you believe to be the final step, start with "So the answer is:".'
     IRCOT_EXAMPLE = "Wikipedia Title: Kurram Garhi\nKurram Garhi is a small village located near the city of Bannu, which is the part of Khyber Pakhtunkhwa province of Pakistan. Its population is approximately 35000. Barren hills are near this village. This village is on the border of Kurram Agency. Other nearby villages are Peppal, Surwangi and Amandi Kala.\n\nWikipedia Title: 2001–02 UEFA Champions League second group stage\nEight winners and eight runners- up from the first group stage were drawn into four groups of four teams, each containing two group winners and two runners- up. Teams from the same country or from the same first round group could not be drawn together. The top two teams in each group advanced to the quarter- finals.\n\nWikipedia Title: Satellite tournament\nA satellite tournament is either a minor tournament or event on a competitive sporting tour or one of a group of such tournaments that form a series played in the same country or region.\n\nWikipedia Title: Trojkrsti\nTrojkrsti is a village in Municipality of Prilep, Republic of Macedonia.\n\nWikipedia Title: Telephone numbers in Ascension Island\nCountry Code:+ 247< br> International Call Prefix: 00 Ascension Island does not share the same country code( +290) with the rest of St Helena.\n\nQuestion: Are both Kurram Garhi and Trojkrsti located in the same country?\nThought: Kurram Garhi is located in the country of Pakistan. Trojkrsti is located in the country of Republic of Macedonia. Thus, they are not in the same country. So the answer is: no.\n\n"
 
-    def __init__(self, config, prompt_template=None, retriever=None, generator=None, max_iter=2):
+    def __init__(self, config, prompt_template=None, retriever=None, generator=None, max_iter=5):
         # if not provide prompt template, use default template provided by IRCOT
         if prompt_template is None:
             prompt_template = PromptTemplate(
@@ -920,60 +920,97 @@ class IRCOTPipeline(BasicPipeline):
 
         self.max_iter = max_iter
 
-    def run_item(self, item):
-        question = item.question
-        retrieval_result, scores = self.retriever.search(question, return_score=True)
-        doc2score = {doc_item["id"]: score for doc_item, score in zip(retrieval_result, scores)}
-        id2doc = {doc_item["id"]: doc_item for doc_item in retrieval_result}
-
-        thoughts = []
+    def run_batch(self, items):
+        # Initialize the necessary data structures
+        batch_thoughts = {item_id: [] for item_id in range(len(items))}
         iter_num = 0
+        batch_retrieval_results = []
+        doc2score_batch = []
+        id2doc_batch = []
+
+        # Initial retrieval for all items in the batch
+        questions = [item.question for item in items]
+        retrieval_results, scoress = self.retriever.batch_search(questions, return_score=True)
+        for retrieval_result, scores in zip(retrieval_results,scoress):   
+            
+            doc2score = {doc_item['id']: score for doc_item, score in zip(retrieval_result, scores)}
+            id2doc = {doc_item['id']: doc_item for doc_item in retrieval_result}
+            batch_retrieval_results.append(retrieval_result)
+            doc2score_batch.append(doc2score)
+            id2doc_batch.append(id2doc)
+
+        # Start the iterative process
+        active_item_ids = list(range(len(items)))  # Track items that need more iterations
         while iter_num < self.max_iter:
-            input_prompt = self.prompt_template.get_string(
-                question=question, retrieval_result=retrieval_result, previous_gen=" ".join(thoughts)
-            )
-            new_thought = self.generator.generate(input_prompt,stop=['.', '\n'])[0]
-            thoughts.append(new_thought)
-            iter_num += 1
-            if "So the answer is:" in new_thought:
-                item.update_output(
-                    f"intermediate_output_iter{iter_num}",
-                    {
-                        "input_prompt": input_prompt,
-                        "new_thought": new_thought,
-                    },
+            # Generate prompts and new thoughts for the active items
+            input_prompts = [
+                self.prompt_template.get_string(
+                    question=items[item_id].question,
+                    retrieval_result=batch_retrieval_results[item_id],
+                    previous_gen=' '.join(batch_thoughts[item_id])
                 )
-                break
+                for item_id in active_item_ids
+            ]
 
-            # retrieve new docs and merge
-            new_retrieval_result, new_scores = self.retriever.search(new_thought, return_score=True)
-            for doc_item, score in zip(new_retrieval_result, new_scores):
-                id2doc[doc_item["id"]] = doc_item
-                doc_id = doc_item["id"]
-                if doc_id in doc2score:
-                    doc2score[doc_id] = max(doc2score[doc_id], score)
+            # Batch generation for active items
+            new_thoughts_batch = self.generator.generate(input_prompts, stop=['.', '\n'])
+            
+            # Update thoughts and determine next active items
+            new_active_item_ids = []
+            for idx, item_id in enumerate(active_item_ids):
+                new_thought = new_thoughts_batch[idx]
+                batch_thoughts[item_id].append(new_thought)
+                
+                # Check for termination condition
+                # Store intermediate outputs
+                if "So the answer is:" in new_thought:
+                    items[item_id].update_output(
+                        f'intermediate_output_iter{iter_num}', 
+                        {
+                            'input_prompt': input_prompts[idx],
+                            'new_thought': new_thought,
+                        },
+                    )
                 else:
-                    doc2score[doc_id] = score
-            sorted_doc_score = sorted(doc2score.items(), key=lambda x: x[1], reverse=False)
-            sorted_doc_id = [t[0] for t in sorted_doc_score]
-            retrieval_result = [id2doc[id] for id in sorted_doc_id]
-
-            item.update_output(
-                f"intermediate_output_iter{iter_num}",
-                {
-                    "input_prompt": input_prompt,
-                    "new_thought": new_thought,
-                    "new_retreival_result": new_retrieval_result,
-                },
-            )
+                    new_active_item_ids.append(item_id)
+
+            # Update active item IDs for the next iteration
+            active_item_ids = new_active_item_ids
+
+            # Perform batch retrieval for new thoughts of active items
+            if active_item_ids:
+                print(len(active_item_ids))
+                print(active_item_ids)
+                new_thoughts_for_retrieval = [batch_thoughts[item_id][-1] for item_id in active_item_ids]
+                new_retrieval_results, new_scoress = self.retriever.batch_search(new_thoughts_for_retrieval, return_score=True)
+
+                for i, item_id in enumerate(active_item_ids):
+                    new_retrieval_result, new_scores = new_retrieval_results[i],new_scoress[i]
+                    
+                    # Update doc2score and id2doc for the current item
+                    for doc_item, score in zip(new_retrieval_result, new_scores):
+                        doc_id = doc_item['id']
+                        id2doc_batch[item_id][doc_id] = doc_item
+                        if doc_id in doc2score_batch[item_id]:
+                            doc2score_batch[item_id][doc_id] = max(doc2score_batch[item_id][doc_id], score)
+                        else:
+                            doc2score_batch[item_id][doc_id] = score
 
-        item.update_output("retrieval_result", retrieval_result)
-        item.update_output("pred", " ".join(thoughts))
-        return item
+                    # Sort and update retrieval results
+                    sorted_doc_score = sorted(doc2score_batch[item_id].items(), key=lambda x: x[1], reverse=False)
+                    sorted_doc_id = [t[0] for t in sorted_doc_score]
+                    batch_retrieval_results[item_id] = [id2doc_batch[item_id][id] for id in sorted_doc_id]
+
+            iter_num += 1
+
+        # Final update for each item in the batch
+        for item_id, item in enumerate(items):
+            item.update_output('retrieval_result', batch_retrieval_results[item_id])
+            item.update_output('pred', ' '.join(batch_thoughts[item_id]))
 
     def run(self, dataset, do_eval=True, pred_process_fun=ircot_pred_parse):
-        for item in tqdm(dataset, desc="Inference: "):
-            self.run_item(item)
+
+        self.run_batch(dataset)
 
         dataset = self.evaluate(dataset, do_eval=do_eval, pred_process_fun=pred_process_fun)
         return dataset

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/pipeline/pipeline.py

@@ -216,10 +216,7 @@ class AdaptivePipeline(BasicPipeline):
         )
 
         self.multi_hop_pipeline = IRCOTPipeline(
-            config,
-            prompt_template=multi_hop_prompt_template,
-            retriever=retriever,
-            generator=generator,
+            config, prompt_template=multi_hop_prompt_template, retriever=retriever, generator=generator, max_iter=5
         )
 
     def run(self, dataset, do_eval=True, pred_process_fun=None):

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/retriever/retriever.py

@@ -7,9 +7,9 @@ from typing import List, Dict
 import functools
 from tqdm import tqdm
 import faiss
-
+import numpy as np
 from flashrag.utils import get_reranker
-from flashrag.retriever.utils import load_corpus, load_docs
+from flashrag.retriever.utils import load_corpus, load_docs, convert_numpy
 from flashrag.retriever.encoder import Encoder, STEncoder
 
 
@@ -136,8 +136,13 @@ class BaseRetriever:
                 self.cache = json.load(f)
 
     def _save_cache(self):
+        self.cache = convert_numpy(self.cache)
+        def custom_serializer(obj):
+            if isinstance(obj, np.float32):  
+                return float(obj)           
+            raise TypeError(f"Type {type(obj)} not serializable")
         with open(self.cache_save_path, "w") as f:
-            json.dump(self.cache, f, indent=4)
+            json.dump(self.cache, f, indent=4, default=custom_serializer)
 
     def _search(self, query: str, num: int, return_score: bool) -> List[Dict[str, str]]:
         r"""Retrieve topk relevant documents in corpus.

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119}/flashrag/retriever/utils.py

@@ -1,8 +1,25 @@
 import json
 import warnings
+from typing import Dict, Any, Union, List, Dict
+import numpy as np
 import datasets
 from transformers import AutoTokenizer, AutoModel, AutoConfig
 
+def convert_numpy(obj: Union[Dict, list, np.ndarray, np.generic]) -> Any:
+    """Recursively convert numpy objects in nested dictionaries or lists to native Python types."""
+    if isinstance(obj, dict):
+        return {k: convert_numpy(v) for k, v in obj.items()}
+    elif isinstance(obj, list):
+        return [convert_numpy(i) for i in obj]
+    elif isinstance(obj, np.ndarray):
+        return obj.tolist()  # Convert numpy arrays to lists
+    elif isinstance(obj, (np.integer, np.floating)):
+        return obj.item()  # Convert numpy scalars to native Python scalars
+    elif isinstance(obj, np.float32):
+        return float(obj)
+    else:
+        return obj  # Return the object as-is if it's neither a dict, list, nor numpy type
+
 
 def load_model(model_path: str, use_fp16: bool = False):
     model_config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
@@ -27,6 +44,7 @@ def pooling(pooler_output, last_hidden_state, attention_mask=None, pooling_metho
     else:
         raise NotImplementedError("Pooling method not implemented!")
 
+
 def set_default_instruction(model_name, is_query=True, is_zh=False):
     instruction = ""
     if "e5" in model_name.lower():
@@ -37,7 +55,7 @@ def set_default_instruction(model_name, is_query=True, is_zh=False):
 
     if "bge" in model_name.lower():
         if is_query:
-            if 'zh' in model_name.lower() or is_zh:
+            if "zh" in model_name.lower() or is_zh:
                 instruction = "为这个句子生成表示以用于检索相关文章："
             else:
                 instruction = "Represent this sentence for searching relevant passages: "
@@ -60,6 +78,8 @@ def parse_query(model_name, query_list, instruction=None):
                     zh_char += 1
             except:
                 continue
+        if len(str) == 0:
+            return False
         if zh_char / len(str) > 0.2:
             return True
         else:
@@ -67,7 +87,7 @@ def parse_query(model_name, query_list, instruction=None):
 
     if isinstance(query_list, str):
         query_list = [query_list]
-        
+
     if instruction is not None:
         instruction = instruction.strip() + " "
     else:
@@ -78,6 +98,7 @@ def parse_query(model_name, query_list, instruction=None):
 
     return query_list
 
+
 def load_corpus(corpus_path: str):
     corpus = datasets.load_dataset("json", data_files=corpus_path, split="train")
     return corpus
@@ -98,3 +119,16 @@ def load_docs(corpus, doc_idxs):
     results = [corpus[int(idx)] for idx in doc_idxs]
 
     return results
+
+
+def parse_image(image):
+    from PIL import Image
+
+    if isinstance(image, str):
+        if image.startswith("http"):
+            import requests
+
+            image = Image.open(requests.get(image, stream=True).raw)
+        else:
+            image = Image.open(image)
+    return image

flashrag-dev-0.1.3.dev20241119/flashrag/version.py

@@ -0,0 +1 @@
+__version__ = "0.1.3dev20241119"

{flashrag-dev-0.1.3.dev20241115 → flashrag-dev-0.1.3.dev20241119/flashrag_dev.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: flashrag-dev
-Version: 0.1.3.dev20241115
+Version: 0.1.3.dev20241119
 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-0.1.3.dev20241115/flashrag/version.py

@@ -1 +0,0 @@
-__version__ = "0.1.3dev20241115"