File size: 8,150 Bytes
1286e81
78209bc
55f46c1
 
c5586ab
12d3e1a
 
 
 
 
 
 
dc376b6
12d3e1a
1286e81
 
 
 
 
 
 
dc376b6
 
d07865c
1286e81
dc376b6
55f46c1
dc376b6
 
12d3e1a
1286e81
 
 
 
 
 
 
 
3d2062e
1286e81
 
cb23311
12d3e1a
 
1286e81
 
 
3d2062e
1286e81
 
12d3e1a
1286e81
dc376b6
 
 
 
 
 
 
1286e81
12d3e1a
1286e81
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dc376b6
1286e81
 
 
 
 
dc376b6
 
 
 
 
 
 
 
 
 
 
 
 
 
1286e81
 
dc376b6
 
 
1286e81
dc376b6
 
 
 
 
 
 
 
 
1286e81
 
dc376b6
1286e81
dc376b6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1286e81
e1d2a79
3d2062e
 
 
1286e81
3d2062e
 
 
1286e81
3d2062e
ecd9808
cb23311
12d3e1a
e1d2a79
1286e81
 
e1d2a79
78209bc
 
 
dc376b6
3d2062e
78209bc
 
 
1286e81
 
e1d2a79
1286e81
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import os
from typing import List, Dict, Tuple, Optional, cast

from pydantic import SecretStr
from _utils.LLMs.LLM_class import LLM
from _utils.vector_stores.Vector_store_class import VectorStore
from setup.easy_imports import (
    Chroma,
    ChatOpenAI,
    PromptTemplate,
    BM25Okapi,
    Response,
    HuggingFaceEmbeddings,
)
import logging
from _utils.gerar_relatorio_modelo_usuario.DocumentSummarizer_simples import (
    DocumentSummarizer,
)
from _utils.models.gerar_relatorio import (
    RetrievalConfig,
)
from cohere import Client
from _utils.splitters.Splitter_class import Splitter


class GerarDocumento:
    openai_api_key = os.environ.get("OPENAI_API_KEY", "")
    cohere_api_key = os.environ.get("COHERE_API_KEY", "")
    resumo_gerado = ""

    def __init__(
        self,
        config: RetrievalConfig,
        embedding_model,
        chunk_size,
        chunk_overlap,
        num_k_rerank,
        model_cohere_rerank,
        # prompt_auxiliar,
        gpt_model,
        gpt_temperature,
        # id_modelo_do_usuario,
        prompt_gerar_documento,
        reciprocal_rank_fusion,
    ):
        self.config = config
        self.logger = logging.getLogger(__name__)
        # self.prompt_auxiliar = prompt_auxiliar
        self.gpt_model = gpt_model
        self.gpt_temperature = gpt_temperature
        self.prompt_gerar_documento = prompt_gerar_documento
        self.reciprocal_rank_fusion = reciprocal_rank_fusion

        self.openai_api_key = self.openai_api_key
        self.cohere_client = Client(self.cohere_api_key)
        self.embeddings = HuggingFaceEmbeddings(model_name=embedding_model)
        self.num_k_rerank = num_k_rerank
        self.model_cohere_rerank = model_cohere_rerank
        self.splitter = Splitter(chunk_size, chunk_overlap)

        self.vector_store = VectorStore(embedding_model)

    def retrieve_with_rank_fusion(
        self, vector_store: Chroma, bm25: BM25Okapi, chunk_ids: List[str], query: str
    ) -> List[Dict]:
        """Combine embedding and BM25 retrieval results"""
        try:
            # Get embedding results
            embedding_results = vector_store.similarity_search_with_score(
                query, k=self.config.num_chunks
            )

            # Convert embedding results to list of (chunk_id, score)
            embedding_list = [
                (doc.metadata["chunk_id"], 1 / (1 + score))
                for doc, score in embedding_results
            ]

            # Get BM25 results
            tokenized_query = query.split()
            bm25_scores = bm25.get_scores(tokenized_query)

            # Convert BM25 scores to list of (chunk_id, score)
            bm25_list = [
                (chunk_ids[i], float(score)) for i, score in enumerate(bm25_scores)
            ]

            # Sort bm25_list by score in descending order and limit to top N results
            bm25_list = sorted(bm25_list, key=lambda x: x[1], reverse=True)[
                : self.config.num_chunks
            ]

            # Normalize BM25 scores
            calculo_max = max(
                [score for _, score in bm25_list]
            )  # Criei este max() pois em alguns momentos estava vindo valores 0, e reclamava que não podia dividir por 0
            max_bm25 = calculo_max if bm25_list and calculo_max else 1
            bm25_list = [(doc_id, score / max_bm25) for doc_id, score in bm25_list]

            # Pass the lists to rank fusion
            result_lists = [embedding_list, bm25_list]
            weights = [self.config.embedding_weight, self.config.bm25_weight]

            combined_results = self.reciprocal_rank_fusion(
                result_lists, weights=weights
            )

            return combined_results

        except Exception as e:
            self.logger.error(f"Error in rank fusion retrieval: {str(e)}")
            raise

    def rank_fusion_get_top_results(
        self,
        vector_store: Chroma,
        bm25: BM25Okapi,
        chunk_ids: List[str],
        query: str = "Summarize the main points of this document",
    ):
        # Get combined results using rank fusion
        ranked_results = self.retrieve_with_rank_fusion(
            vector_store, bm25, chunk_ids, query
        )

        # Prepare context and track sources
        contexts = []
        sources = []

        # Get full documents for top results
        for chunk_id, score in ranked_results[: self.config.num_chunks]:
            results = vector_store.get(
                where={"chunk_id": chunk_id}, include=["documents", "metadatas"]
            )

            if results["documents"]:
                context = results["documents"][0]
                metadata = results["metadatas"][0]

                contexts.append(context)
                sources.append(
                    {
                        "content": context,
                        "page": metadata["page"],
                        "chunk_id": chunk_id,
                        "relevance_score": score,
                        "context": metadata.get("context", ""),
                    }
                )

        return sources, contexts

    def select_model_for_last_requests(self, llm_ultimas_requests: str):
        llm_instance = LLM()
        if llm_ultimas_requests == "gpt-4o-mini":
            llm = ChatOpenAI(
                temperature=self.gpt_temperature,
                model=self.gpt_model,
                api_key=SecretStr(self.openai_api_key),
            )
        elif llm_ultimas_requests == "deepseek-chat":
            llm = llm_instance.deepseek()
        elif llm_ultimas_requests == "gemini-2.0-flash":
            llm = llm_instance.google_gemini("gemini-2.0-flash")
        return llm

    async def gerar_documento_final(
        self,
        vector_store: Chroma,
        bm25: BM25Okapi,
        chunk_ids: List[str],
        llm_ultimas_requests: str,
        query: str = "Summarize the main points of this document",
    ) -> List[Dict]:
        try:
            sources, contexts = self.rank_fusion_get_top_results(
                vector_store, bm25, chunk_ids, query
            )

            llm = self.select_model_for_last_requests(llm_ultimas_requests)
            # prompt_auxiliar = PromptTemplate(
            #     template=self.prompt_auxiliar, input_variables=["context"]
            # )

            # resumo_auxiliar_do_documento = llm.invoke(
            #     prompt_auxiliar.format(context="\n\n".join(contexts))
            # )

            # self.resumo_gerado = cast(str, resumo_auxiliar_do_documento.content)

            prompt_gerar_documento = PromptTemplate(
                template=self.prompt_gerar_documento,
                input_variables=["context"],
            )

            documento_gerado = cast(
                str,
                llm.invoke(
                    prompt_gerar_documento.format(
                        context="\n\n".join(contexts),
                        # modelo_usuario=serializer.data["modelo"],
                    )
                ).content,
            )

            # Split the response into paragraphs
            summaries = [p.strip() for p in documento_gerado.split("\n\n") if p.strip()]

            # Create structured output
            structured_output = []
            for idx, summary in enumerate(summaries):
                source_idx = min(idx, len(sources) - 1)
                structured_output.append(
                    {
                        "content": summary,
                        "source": {
                            "page": sources[source_idx]["page"],
                            "text": sources[source_idx]["content"][:200] + "...",
                            "context": sources[source_idx]["context"],
                            "relevance_score": sources[source_idx]["relevance_score"],
                            "chunk_id": sources[source_idx]["chunk_id"],
                        },
                    }
                )

            return structured_output

        except Exception as e:
            self.logger.error(f"Error generating enhanced summary: {str(e)}")
            raise