File size: 11,006 Bytes
7a8b33f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
import faiss
import shutil
from beartype import beartype
import numpy as np
import json
import argparse
from zsvision.zs_utils import BlockTimer
import tiktoken
from pathlib import Path
from langchain.text_splitter import RecursiveCharacterTextSplitter
from pipeline_paths import PIPELINE_PATHS
from llm_api_utils import (
    init_openai_with_api_key,
    EMBEDDING_DIMENSIONS,
    PRICE_PER_1K_TOKENS,
)
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings
from langchain.docstore.in_memory import InMemoryDocstore


class EmbedResults:
    def __init__(
        self,
        embedding_model="ada",
        limit=0,
        refresh=False,
        refresh_faiss_db=False,
        text_embedding_chunk_size=500,
        filter_str="",
    ):
        self.embedding_model = embedding_model
        self.limit = limit
        self.refresh = refresh
        self.refresh_faiss_db = refresh_faiss_db
        self.text_embedding_chunk_size = text_embedding_chunk_size
        self.filter_str = filter_str

    @beartype
    def compute_embeddings_from_chunks(
        self, embedding_function: OpenAIEmbeddings, metadatas: list, faiss_db
    ):
        doc_chunks = []
        metadatas_without_chunks = []
        for metadata in metadatas:
            doc_chunk = metadata.pop("doc_chunk")
            doc_chunks.append(doc_chunk)
            metadatas_without_chunks.append(metadata)

        with BlockTimer(f"Embedding {len(metadatas)} fragments"):
            embeddings = embedding_function.embed_documents(doc_chunks)
            # account for name mangling in Python
            faiss_db._FAISS__add(doc_chunks, embeddings, metadatas_without_chunks)

        return faiss_db

    @beartype
    def parse_date_of_fetching(self, data: dict) -> str:
        evidence_keys = {
            "search_results_fetched",
            "results_fetched_from_wikipedia_1M_with_cohere-22-12",
        }
        for key in evidence_keys:
            if key in data["dates"]:
                evidence_fetched_date = data["dates"][key]
                return evidence_fetched_date
        raise ValueError(f"Could not find evidence fetched date in {data['dates']}")

    def embed_for_uuid(self, srcs):
        init_openai_with_api_key()

        embedding_function = OpenAIEmbeddings()

        index = faiss.IndexFlatL2(EMBEDDING_DIMENSIONS[self.embedding_model])
        docstore = InMemoryDocstore({})
        index_to_docstore_id = {}
        faiss_db = FAISS(
            embedding_function=embedding_function.embed_query,
            index=index,
            docstore=docstore,
            index_to_docstore_id=index_to_docstore_id,
        )

        already_embedded_chunks = {
            doc.metadata["chunk_tag"] for doc in faiss_db.docstore._dict.values()
        }

        splitter = RecursiveCharacterTextSplitter(
            chunk_size=self.text_embedding_chunk_size,
            chunk_overlap=0,
        )

        kwarg_list = []
        seen_links = set()
        metadatas = []
        total_chunks = 0
        chunks_to_embed = 0
        chunks_to_skip = 0

        for data in srcs:
            evidence_fetched_date = self.parse_date_of_fetching(data)

            for document in data["documents"]:
                for search_result in document["search_results"]:
                    # Don't embed the same link twice
                    if search_result["link"] in seen_links:
                        continue
                    seen_links.add(search_result["link"])

                    doc_chunks = [
                        doc.page_content
                        for doc in splitter.create_documents([search_result["text"]])
                    ]
                    chunk_tags = [
                        f"{search_result['link']}-chunk-{idx}-chunk_sz-{self.text_embedding_chunk_size}"
                        for idx in range(len(doc_chunks))
                    ]
                    for doc_chunk, chunk_tag in zip(doc_chunks, chunk_tags):
                        if chunk_tag not in already_embedded_chunks:
                            metadatas.append(
                                {
                                    "doc_chunk": doc_chunk,
                                    "link": search_result["link"],
                                    "chunk_tag": chunk_tag,
                                    "date_accessed": evidence_fetched_date,
                                    "query": document["claim"],
                                }
                            )
                            chunks_to_embed += 1
                        else:
                            chunks_to_skip += 1
                    total_chunks += len(doc_chunks)

        encoding = tiktoken.encoding_for_model(self.embedding_model)
        doc_chunks = [x["doc_chunk"] for x in metadatas]
        num_words = len(" ".join(doc_chunks).split())
        num_tokens = len(encoding.encode("".join(doc_chunks)))

        print(
            f"Created {total_chunks} chunks of text to answer from {len(seen_links)} websites"
        )
        print(
            f"Embedding {chunks_to_embed} (skipping {chunks_to_skip}) chunks of text from {len(seen_links)} websites)"
        )
        print(
            f"Embedding {num_tokens} tokens ({num_words} words) from {len(doc_chunks)} chunks"
        )
        print(
            f"Step5: Estimated cost: {num_tokens * PRICE_PER_1K_TOKENS[self.embedding_model]['embed'] / 1000:.2f} USD"
        )

        if metadatas:
            self.compute_embeddings_from_chunks(
                embedding_function=embedding_function,
                faiss_db=faiss_db,
                metadatas=metadatas,
            )

            return faiss_db
        return None

    def embed(self):
        init_openai_with_api_key()
        src_paths = []
        for evidence_key in (
            "google_search_results_evidence",
            "cohere_wikipedia_evidence",
        ):
            evidence_paths = list(PIPELINE_PATHS[evidence_key].glob("**/*.json"))
            src_paths.extend(evidence_paths)

        if self.filter_str:
            num_paths = len(src_paths)
            src_paths = [
                src_path for src_path in src_paths if self.filter_str in src_path.name
            ]
            print(
                f"Filtering for {self.filter_str} (from {num_paths} to {len(src_paths)})"
            )

        print(f"Found {len(src_paths)} collections of evidence")
        src_paths = sorted(src_paths)

        embedding_function = OpenAIEmbeddings()
        faiss_persist_dir = (
            PIPELINE_PATHS["faiss_db_embeddings_for_evidence"]
            / f"{self.embedding_model}_chunk_size_{self.text_embedding_chunk_size}"
        )

        if faiss_persist_dir.exists():
            if self.refresh_faiss_db:
                print(f"Deleting existing database at {faiss_persist_dir}")
                shutil.rmtree(faiss_persist_dir)

        # check which chunks we've already embedded to avoid duplication
        if faiss_persist_dir.exists() and not self.refresh_faiss_db:
            faiss_db = FAISS.load_local(
                folder_path=str(faiss_persist_dir), embeddings=embedding_function
            )
            print(f"Found existing database at {faiss_persist_dir}, using... ")
        else:
            index = faiss.IndexFlatL2(EMBEDDING_DIMENSIONS[self.embedding_model])
            docstore = InMemoryDocstore({})
            index_to_docstore_id = {}
            faiss_db = FAISS(
                embedding_function=embedding_function.embed_query,
                index=index,
                docstore=docstore,
                index_to_docstore_id=index_to_docstore_id,
            )
            print(f"Persisting intialised database to {faiss_persist_dir}")
            faiss_db.save_local(folder_path=str(faiss_persist_dir))

        already_embedded_chunks = {
            doc.metadata["chunk_tag"] for doc in faiss_db.docstore._dict.values()
        }

        splitter = RecursiveCharacterTextSplitter(
            chunk_size=self.text_embedding_chunk_size,
            chunk_overlap=0,
        )

        kwarg_list = []
        seen_links = set()
        metadatas = []
        total_chunks = 0
        chunks_to_embed = 0
        chunks_to_skip = 0

        for src_path in src_paths:
            with open(src_path, "r") as f:
                data = json.load(f)

            evidence_fetched_date = self.parse_date_of_fetching(data)

            for document in data["documents"]:
                for search_result in document["search_results"]:
                    # Don't embed the same link twice
                    if search_result["link"] in seen_links:
                        continue
                    seen_links.add(search_result["link"])

                    doc_chunks = [
                        doc.page_content
                        for doc in splitter.create_documents([search_result["text"]])
                    ]
                    chunk_tags = [
                        f"{search_result['link']}-chunk-{idx}-chunk_sz-{self.text_embedding_chunk_size}"
                        for idx in range(len(doc_chunks))
                    ]
                    for doc_chunk, chunk_tag in zip(doc_chunks, chunk_tags):
                        if chunk_tag not in already_embedded_chunks:
                            metadatas.append(
                                {
                                    "doc_chunk": doc_chunk,
                                    "link": search_result["link"],
                                    "chunk_tag": chunk_tag,
                                    "date_accessed": evidence_fetched_date,
                                    "query": document["claim"],
                                }
                            )
                            chunks_to_embed += 1
                        else:
                            chunks_to_skip += 1
                    total_chunks += len(doc_chunks)

        encoding = tiktoken.encoding_for_model(self.embedding_model)
        doc_chunks = [x["doc_chunk"] for x in metadatas]
        num_words = len(" ".join(doc_chunks).split())
        num_tokens = len(encoding.encode("".join(doc_chunks)))

        print(
            f"Created {total_chunks} chunks of text to answer from {len(seen_links)} websites"
        )
        print(
            f"Embedding {chunks_to_embed} (skipping {chunks_to_skip}) chunks of text from {len(seen_links)} websites)"
        )
        print(
            f"Embedding {num_tokens} tokens ({num_words} words) from {len(doc_chunks)} chunks"
        )
        print(
            f"Estimated cost: {num_tokens * PRICE_PER_1K_TOKENS[self.embedding_model]['embed'] / 1000:.2f} USD"
        )

        if metadatas:
            self.compute_embeddings_from_chunks(
                embedding_function=embedding_function,
                faiss_persist_dir=faiss_persist_dir,
                metadatas=metadatas,
            )