File size: 4,140 Bytes
481f3b1
 
 
 
 
7335378
 
481f3b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
088e816
481f3b1
 
 
6b43c86
481f3b1
 
 
088e816
 
 
 
481f3b1
088e816
481f3b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5664fc8
 
 
 
481f3b1
 
 
 
088e816
 
 
 
 
 
 
 
481f3b1
 
 
 
 
 
 
 
 
 
 
 
 
088e816
481f3b1
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
import sys
import os
from contextlib import contextmanager

from ..reranker import rerank_docs
from ...knowledge.retriever import ClimateQARetriever




def divide_into_parts(target, parts):
    # Base value for each part
    base = target // parts
    # Remainder to distribute
    remainder = target % parts
    # List to hold the result
    result = []
    
    for i in range(parts):
        if i < remainder:
            # These parts get base value + 1
            result.append(base + 1)
        else:
            # The rest get the base value
            result.append(base)
    
    return result


@contextmanager
def suppress_output():
    # Open a null device
    with open(os.devnull, 'w') as devnull:
        # Store the original stdout and stderr
        old_stdout = sys.stdout
        old_stderr = sys.stderr
        # Redirect stdout and stderr to the null device
        sys.stdout = devnull
        sys.stderr = devnull
        try:
            yield
        finally:
            # Restore stdout and stderr
            sys.stdout = old_stdout
            sys.stderr = old_stderr



def make_retriever_node(vectorstore,reranker,rerank_by_question=True, k_final=15, k_before_reranking=100, k_summary=5):

    def retrieve_documents(state):
        
        POSSIBLE_SOURCES = ["IPCC","IPBES","IPOS"] # ,"OpenAlex"]
        questions = state["questions"]
        
        # Use sources from the user input or from the LLM detection
        if "sources_input" not in state or state["sources_input"] is None:
            sources_input = ["auto"]
        else:
            sources_input = state["sources_input"]
        auto_mode = "auto" in sources_input

        # There are several options to get the final top k
        # Option 1 - Get 100 documents by question and rerank by question
        # Option 2 - Get 100/n documents by question and rerank the total
        if rerank_by_question:
            k_by_question = divide_into_parts(k_final,len(questions))
        
        docs = []
        
        for i,q in enumerate(questions):
            
            sources = q["sources"]
            question = q["question"]
            
            # If auto mode, we use the sources detected by the LLM
            if auto_mode:
                sources = [x for x in sources if x in POSSIBLE_SOURCES]
                
            # Otherwise, we use the config
            else:
                sources = sources_input
                
            # Search the document store using the retriever
            # Configure high top k for further reranking step
            retriever = ClimateQARetriever(
                vectorstore=vectorstore,
                sources = sources,
                # reports = ias_reports,
                min_size = 200,
                k_summary = k_summary,
                k_total = k_before_reranking,
                threshold = 0.5,
            )
            docs_question = retriever.get_relevant_documents(question)
            
            # Rerank
            if reranker is not None:
                with suppress_output():
                    docs_question = rerank_docs(reranker,docs_question,question)
            else:
                # Add a default reranking score
                for doc in docs_question:
                    doc.metadata["reranking_score"] = doc.metadata["similarity_score"]
                
            # If rerank by question we select the top documents for each question
            if rerank_by_question:
                docs_question = docs_question[:k_by_question[i]]
                
            # Add sources used in the metadata
            for doc in docs_question:
                doc.metadata["sources_used"] = sources
            
            # Add to the list of docs
            docs.extend(docs_question)
            
        # Sorting the list in descending order by rerank_score
        # Then select the top k
        docs = sorted(docs, key=lambda x: x.metadata["reranking_score"], reverse=True)
        docs = docs[:k_final]
        
        new_state = {"documents":docs}
        return new_state
    
    return retrieve_documents