Skip to content

Vector Store Utilities

  • Code for loading data into the vector store.

What to look for

  • DataLoader: If you want to modify chunking
  • DocumentProcessor: If you want to modify how the unique documents are obtained and/or add other methods
  • VectorStoreManager: If you want to modify how the documents are embedded and loaded to the vector store

DataLoader

Description: Used to chunk data

Source code in backend/modules/vector_store_utils.py 
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
class DataLoader:
    """
    Description: Used to chunk data
    """

    def __init__(
        self,
        metadata_df: pd.DataFrame,
        page_content_column: str,
        chunk_size: int = 1000,
        chunk_overlap: int = 150,
    ):
        self.metadata_df = metadata_df
        self.page_content_column = page_content_column
        self.chunk_size = chunk_size
        self.chunk_overlap = (
            chunk_overlap if self.chunk_size > chunk_overlap else self.chunk_size
        )

    def load_and_process_data(self) -> list:
        """
        Description: Recursively chunk data before embedding
        """
        loader = DataFrameLoader(
            self.metadata_df, page_content_column=self.page_content_column
        )
        documents = loader.load()

        text_splitter = RecursiveCharacterTextSplitter(
            chunk_size=self.chunk_size, chunk_overlap=self.chunk_overlap
        )
        documents = text_splitter.split_documents(documents)

        return documents

load_and_process_data()

Description: Recursively chunk data before embedding

Source code in backend/modules/vector_store_utils.py 
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
def load_and_process_data(self) -> list:
    """
    Description: Recursively chunk data before embedding
    """
    loader = DataFrameLoader(
        self.metadata_df, page_content_column=self.page_content_column
    )
    documents = loader.load()

    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=self.chunk_size, chunk_overlap=self.chunk_overlap
    )
    documents = text_splitter.split_documents(documents)

    return documents

DocumentProcessor

Description: Used to generate unique documents based on text content to prevent duplicates during embedding

Source code in backend/modules/vector_store_utils.py 
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
class DocumentProcessor:
    """
    Description: Used to generate unique documents based on text content to prevent duplicates during embedding
    """

    @staticmethod
    def generate_unique_documents(documents: list, db: Chroma) -> tuple:
        """
        Description: Sometimes the text content of the data is the same, this ensures that does not happen by computing a string matching
        """
        new_document_ids = set([str(x.metadata["did"]) for x in documents])
        print(f"[INFO] Generating unique documents. Total documents: {len(documents)}")
        try:
            old_dids = set([str(x["did"]) for x in db.get()["metadatas"]])
        except KeyError:
            old_dids = set([str(x["id"]) for x in db.get()["metadatas"]])

        new_dids = new_document_ids - old_dids
        documents = [x for x in documents if str(x.metadata["did"]) in new_dids]
        ids = [
            str(uuid.uuid5(uuid.NAMESPACE_DNS, doc.page_content)) for doc in documents
        ]

        unique_ids = list(set(ids))
        seen_ids = set()
        unique_docs = [
            doc
            for doc, id in zip(documents, ids)
            if id not in seen_ids and (seen_ids.add(id) or True)
        ]

        return unique_docs, unique_ids

generate_unique_documents(documents, db) staticmethod

Description: Sometimes the text content of the data is the same, this ensures that does not happen by computing a string matching

Source code in backend/modules/vector_store_utils.py 
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
@staticmethod
def generate_unique_documents(documents: list, db: Chroma) -> tuple:
    """
    Description: Sometimes the text content of the data is the same, this ensures that does not happen by computing a string matching
    """
    new_document_ids = set([str(x.metadata["did"]) for x in documents])
    print(f"[INFO] Generating unique documents. Total documents: {len(documents)}")
    try:
        old_dids = set([str(x["did"]) for x in db.get()["metadatas"]])
    except KeyError:
        old_dids = set([str(x["id"]) for x in db.get()["metadatas"]])

    new_dids = new_document_ids - old_dids
    documents = [x for x in documents if str(x.metadata["did"]) in new_dids]
    ids = [
        str(uuid.uuid5(uuid.NAMESPACE_DNS, doc.page_content)) for doc in documents
    ]

    unique_ids = list(set(ids))
    seen_ids = set()
    unique_docs = [
        doc
        for doc, id in zip(documents, ids)
        if id not in seen_ids and (seen_ids.add(id) or True)
    ]

    return unique_docs, unique_ids

VectorStoreManager

Description: Manages the Vector store (chromadb) and takes care of data ingestion, loading the embedding model and embedding the data before adding it to the vector store

Source code in backend/modules/vector_store_utils.py 
 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
class VectorStoreManager:
    """
    Description: Manages the Vector store (chromadb) and takes care of data ingestion, loading the embedding model and embedding the data before adding it to the vector store
    """

    def __init__(self, chroma_client: ClientAPI, config: dict):
        self.chroma_client = chroma_client
        self.config = config
        self.chunk_size = 100

    def chunk_dataframe(self, df, chunk_size):
        """
        Description: Chunk dataframe for use with chroma metadata saving
        """
        for i in range(0, df.shape[0], self.chunk_size):
            yield df.iloc[i : i + self.chunk_size]

    def add_df_chunks_to_db(self, metadata_df):
        """
        Description: Add chunks from a dataframe for use with chroma metadata saving
        """
        collec = self.chroma_client.get_or_create_collection("metadata")
        for chunk in tqdm(
            self.chunk_dataframe(metadata_df, self.chunk_size),
            total=(len(metadata_df) // self.chunk_size) + 1,
        ):
            ids = chunk["did"].astype(str).tolist()
            documents = chunk["description"].astype(str).tolist()
            metadatas = chunk.to_dict(orient="records")

            # Add to ChromaDB collection
            collec.add(ids=ids, documents=documents, metadatas=metadatas)

    def load_model(self) -> HuggingFaceEmbeddings | None:
        """
        Description: Load a model from Hugging face for embedding
        """
        print("[INFO] Loading model...")
        model_kwargs = {"device": self.config["device"], "trust_remote_code": True}
        encode_kwargs = {"normalize_embeddings": True}
        embeddings = HuggingFaceEmbeddings(
            model_name=self.config["embedding_model"],
            model_kwargs=model_kwargs,
            encode_kwargs=encode_kwargs,
            show_progress=False,
        )
        print("[INFO] Model loaded.")
        return embeddings

    def get_collection_name(self) -> str:
        """
        Description: Fixes some collection names. (workaround from OpenML API)
        """
        return {"dataset": "datasets", "flow": "flows"}.get(
            self.config["type_of_data"], "default"
        )

    def load_vector_store(
        self, embeddings: HuggingFaceEmbeddings, collection_name: str
    ) -> Chroma:
        """
        Description: Persist directory. If does not exist, cannot be served
        """
        if not os.path.exists(self.config["persist_dir"]):
            raise Exception(
                "Persist directory does not exist. Please run the training pipeline first."
            )

        return Chroma(
            client=self.chroma_client,
            persist_directory=self.config["persist_dir"],
            embedding_function=embeddings,
            collection_name=collection_name,
        )

    @staticmethod
    def add_documents_to_db(db, unique_docs, unique_ids, bs=512):
        """
        Description: Add documents to Chroma DB in batches of bs
        """
        if len(unique_docs) < bs:
            db.add_documents(unique_docs, ids=unique_ids)
        else:
            for i in range(0, len(unique_docs), bs):
                db.add_documents(unique_docs[i : i + bs], ids=unique_ids[i : i + bs])

    def create_vector_store(self, metadata_df: pd.DataFrame) -> Chroma:
        """
        Description: Load embeddings, get chunked data, subset if needed , find unique, and then finally add to ChromaDB
        """
        embeddings = self.load_model()
        collection_name = self.get_collection_name()

        db = Chroma(
            client=self.chroma_client,
            embedding_function=embeddings,
            persist_directory=self.config["persist_dir"],
            collection_name=collection_name,
        )

        data_loader = DataLoader(
            metadata_df,
            page_content_column="Combined_information",
            chunk_size=self.config["chunk_size"],
        )
        documents = data_loader.load_and_process_data()

        if self.config["testing_flag"]:
            if self.config["test_subset"]:
                print("[INFO] Subsetting the data.")
                documents = documents[:500]

        unique_docs, unique_ids = DocumentProcessor.generate_unique_documents(
            documents, db
        )
        print(
            f"Number of unique documents: {len(unique_docs)} vs Total documents: {len(documents)}"
        )

        if len(unique_docs) == 0:
            print("No new documents to add.")
        else:
            self.add_documents_to_db(db, unique_docs, unique_ids)

        return db

add_df_chunks_to_db(metadata_df)

Description: Add chunks from a dataframe for use with chroma metadata saving

Source code in backend/modules/vector_store_utils.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
def add_df_chunks_to_db(self, metadata_df):
    """
    Description: Add chunks from a dataframe for use with chroma metadata saving
    """
    collec = self.chroma_client.get_or_create_collection("metadata")
    for chunk in tqdm(
        self.chunk_dataframe(metadata_df, self.chunk_size),
        total=(len(metadata_df) // self.chunk_size) + 1,
    ):
        ids = chunk["did"].astype(str).tolist()
        documents = chunk["description"].astype(str).tolist()
        metadatas = chunk.to_dict(orient="records")

        # Add to ChromaDB collection
        collec.add(ids=ids, documents=documents, metadatas=metadatas)

add_documents_to_db(db, unique_docs, unique_ids, bs=512) staticmethod

Description: Add documents to Chroma DB in batches of bs

Source code in backend/modules/vector_store_utils.py 
158
159
160
161
162
163
164
165
166
167
@staticmethod
def add_documents_to_db(db, unique_docs, unique_ids, bs=512):
    """
    Description: Add documents to Chroma DB in batches of bs
    """
    if len(unique_docs) < bs:
        db.add_documents(unique_docs, ids=unique_ids)
    else:
        for i in range(0, len(unique_docs), bs):
            db.add_documents(unique_docs[i : i + bs], ids=unique_ids[i : i + bs])

chunk_dataframe(df, chunk_size)

Description: Chunk dataframe for use with chroma metadata saving

Source code in backend/modules/vector_store_utils.py 
93
94
95
96
97
98
def chunk_dataframe(self, df, chunk_size):
    """
    Description: Chunk dataframe for use with chroma metadata saving
    """
    for i in range(0, df.shape[0], self.chunk_size):
        yield df.iloc[i : i + self.chunk_size]

create_vector_store(metadata_df)

Description: Load embeddings, get chunked data, subset if needed , find unique, and then finally add to ChromaDB

Source code in backend/modules/vector_store_utils.py 
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
def create_vector_store(self, metadata_df: pd.DataFrame) -> Chroma:
    """
    Description: Load embeddings, get chunked data, subset if needed , find unique, and then finally add to ChromaDB
    """
    embeddings = self.load_model()
    collection_name = self.get_collection_name()

    db = Chroma(
        client=self.chroma_client,
        embedding_function=embeddings,
        persist_directory=self.config["persist_dir"],
        collection_name=collection_name,
    )

    data_loader = DataLoader(
        metadata_df,
        page_content_column="Combined_information",
        chunk_size=self.config["chunk_size"],
    )
    documents = data_loader.load_and_process_data()

    if self.config["testing_flag"]:
        if self.config["test_subset"]:
            print("[INFO] Subsetting the data.")
            documents = documents[:500]

    unique_docs, unique_ids = DocumentProcessor.generate_unique_documents(
        documents, db
    )
    print(
        f"Number of unique documents: {len(unique_docs)} vs Total documents: {len(documents)}"
    )

    if len(unique_docs) == 0:
        print("No new documents to add.")
    else:
        self.add_documents_to_db(db, unique_docs, unique_ids)

    return db

get_collection_name()

Description: Fixes some collection names. (workaround from OpenML API)

Source code in backend/modules/vector_store_utils.py 
132
133
134
135
136
137
138
def get_collection_name(self) -> str:
    """
    Description: Fixes some collection names. (workaround from OpenML API)
    """
    return {"dataset": "datasets", "flow": "flows"}.get(
        self.config["type_of_data"], "default"
    )

load_model()

Description: Load a model from Hugging face for embedding

Source code in backend/modules/vector_store_utils.py 
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
def load_model(self) -> HuggingFaceEmbeddings | None:
    """
    Description: Load a model from Hugging face for embedding
    """
    print("[INFO] Loading model...")
    model_kwargs = {"device": self.config["device"], "trust_remote_code": True}
    encode_kwargs = {"normalize_embeddings": True}
    embeddings = HuggingFaceEmbeddings(
        model_name=self.config["embedding_model"],
        model_kwargs=model_kwargs,
        encode_kwargs=encode_kwargs,
        show_progress=False,
    )
    print("[INFO] Model loaded.")
    return embeddings

load_vector_store(embeddings, collection_name)

Description: Persist directory. If does not exist, cannot be served

Source code in backend/modules/vector_store_utils.py 
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def load_vector_store(
    self, embeddings: HuggingFaceEmbeddings, collection_name: str
) -> Chroma:
    """
    Description: Persist directory. If does not exist, cannot be served
    """
    if not os.path.exists(self.config["persist_dir"]):
        raise Exception(
            "Persist directory does not exist. Please run the training pipeline first."
        )

    return Chroma(
        client=self.chroma_client,
        persist_directory=self.config["persist_dir"],
        embedding_function=embeddings,
        collection_name=collection_name,
    )