📚 PreVectorChunks

A lightweight utility for document chunking and vector database upserts — designed for developers building RAG (Retrieval-Augmented Generation) solutions.

---

✨ Who Needs This Module?

Any developer working with:

• RAG pipelines
• Vector Databases (like Pinecone, Weaviate, etc.)
• AI applications requiring similar content retrieval

---

🎯 What Does This Module Do?

This module helps you:

• Chunk documents into smaller fragments using:
  • a pretrained Reinforcement Learning based model or
  • a pretrained Reinforcement Learning based model with proposition indexing or
  • standard word chunking
  • recursive character based chunking
  • character based chunking
• Insert (upsert) fragments into a vector database
• Fetch & update existing chunks from a vector database

---

📦 Installation

pip install prevectorchunks-core

How to import in a file:

from PreVectorChunks.services import chunk_documents_crud_vdb

Use .env for API keys:IMPORTANT: PLEASE ENSURE TO PROVIDE YOUR OPENAI_API_KEY as MINIMUM in an .env file or as required

PINECONE_API_KEY=YOUR_API_KEY
OPENAI_API_KEY=YOUR_API_KEY

---

📄 Functions

1. chunk_documents

chunk_documents(instructions, file_path="content_playground/content.json", splitter_config=SplitterConfig())

Splits the content of a document into smaller, manageable chunks. - Five types of document chunking

• Chunking using Reinforcement Learning based pretrained model +(enable/disable LLM to structure the chunked text - default is enabled)
• Chunking using Reinforcement Learning based pretrained model and proposition indexing +(enable/disable LLM to structure the chunked text - default is enabled)
• Recursive Character based chunking +(enable/disable LLM to structure the chunked text - default is enabled)
• Standard word based chunking+(enable/disable LLM to structure the chunked text - default is enabled)
• Simple character based chunking +(enable/disable LLM to structure the chunked text - default is enabled)

Parameters

• instructions (dict or str): Additional rules or guidance for how the document should be split.
  • Example: "split my content by biggest headings"
• file_path (str): Binary file or file path to the input file containing the content or content of the file. Default: "content_playground/content.json".
• splitter_config (optional) (SplitterConfig): (if none provided standard split takes place) Object that defines chunking behavior, e.g., chunk_size, chunk_overlap, separator, split_type.
• i.e. splitter_config = SplitterConfig(chunk_size= 300, chunk_overlap= 0,separators=["\n"],split_type=SplitType.RECURSIVE.value)
• (chunk_size refers to size in characters (i.e. 100 characters) when RECURSIVE is used)
• i.e. splitter_config = SplitterConfig(chunk_size= 300, chunk_overlap= 0,separators=["\n"],split_type=SplitType.CHARACTER.value)
• • (chunk_size refers to size in characters (i.e. 100 characters) when CHARACTER is used)
• i.e. splitter_config = SplitterConfig(chunk_size= 300, chunk_overlap= 0,separators=["\n"],split_type=SplitType.STANDARD.value)
• • (chunk_size refers to size in words (i.e. 100 characters) when STANDARD is used)
• i.e. splitter_config = SplitterConfig(separators=["\n"], split_type=SplitType.R_PRETRAINED.value, min_rl_chunk_size=5, max_rl_chunk_size=50,enableLLMTouchUp=False)
• • (min_rl_chunk_size and max_rl_chunk_size refers to size in sentences (i.e. 100 sentences) when R_PRETRAINED is used)
• i.e. splitter_config = SplitterConfig(separators=["\n"], split_type=SplitType.R_PRETRAINED_PROPOSITION.value, min_rl_chunk_size=5, max_rl_chunk_size=50,enableLLMTouchUp=False)
• • (min_rl_chunk_size and max_rl_chunk_size refers to size in sentences (i.e. 100 sentences) when R_PRETRAINED_PROPOSITION is used) Returns
• A list of chunked strings including a unique id, a meaningful title and chunked text

Use Cases

• Preparing text for LLM ingestion
• Splitting text by structure (headings, paragraphs)
• Vector database indexing

---

2. chunk_and_upsert_to_vdb

chunk_and_upsert_to_vdb(index_n, instructions, file_path="content_playground/content.json", splitter_config=SplitterConfig())

Splits a document into chunks (via chunk_documents) and inserts them into a Vector Database.

Parameters

• index_n (str): The name of the VDB index where chunks should be stored.
• instructions (dict or str): Rules for splitting content (same as chunk_documents).
• file_path (str): Path to the document file or content of the file. Default: "content_playground/content.json".
• splitter_config (SplitterConfig): Object that defines chunking behavior.

Returns

• Confirmation of successful insert into the VDB.

Use Cases

• Automated document preprocessing and storage for vector search
• Preparing embeddings for semantic search

---

3. fetch_vdb_chunks_grouped_by_document_name

fetch_vdb_chunks_grouped_by_document_name(index_n)

Fetches existing chunks stored in the Vector Database, grouped by document name.

Parameters

• index_n (str): The name of the VDB index.

Returns

• A dictionary or list of chunks grouped by document name.

Use Cases

• Retrieving all chunks of a specific document
• Verifying what content has been ingested into the VDB

---

4. update_vdb_chunks_grouped_by_document_name

update_vdb_chunks_grouped_by_document_name(index_n, dataset)

Updates existing chunks in the Vector Database by document name.

Parameters

• index_n (str): The name of the VDB index.
• dataset (dict or list): The new data (chunks) to update existing entries.

Returns

• Confirmation of update status.

Use Cases

• Keeping VDB chunks up to date when documents change
• Re-ingesting revised or corrected content

---

5. markdown_and_chunk_documents

from prevectorchunks_core.services.markdown_and_chunk_documents import MarkdownAndChunkDocuments

markdown_processor = MarkdownAndChunkDocuments()
mapped_chunks = markdown_processor.markdown_and_chunk_documents("example.pdf")

Description
This new function automatically:

1. Converts a document (PDF, DOCX, etc.) into images using DocuToImageConverter.
2. Extracts Markdown and text content from those images using DocuToMarkdownExtractor (powered by GPT).
3. Converts the extracted markdown text into RL-based chunks using ChunkMapper and chunk_documents.
4. Merges unmatched markdown segments into the final structured output.

Parameters

• file_path (str): Path to the document (PDF, DOCX, or image) you want to process.

Returns

• mapped_chunks (list[dict]): A list of markdown-based chunks with both markdown and chunked text content.

Example

if __name__ == "__main__":
    markdown_processor = MarkdownAndChunkDocuments()
    mapped_chunks = markdown_processor.markdown_and_chunk_documents("421307-nz-au-top-loading-washer-guide-shorter.pdf")
    print(mapped_chunks)

Use Cases

• End-to-end document-to-markdown-to-chunks pipeline
• Automating preprocessing for RAG/LLM ingestion
• Extracting structured markdown for semantic search or content indexing

---

🚀 Example Workflow

from prevectorchunks_core.config import SplitterConfig

splitter_config = SplitterConfig(chunk_size=150, chunk_overlap=0, separator=["\n"], split_type=SplitType.R_PRETRAINED_PROPOSITION.value)

# Step 1: Chunk a document
chunks = chunk_documents(
    instructions="split my content by biggest headings",
    file_path="content_playground/content.json",
    splitter_config=splitter_config
)

splitter_config = SplitterConfig(chunk_size=300, chunk_overlap=0, separators=["\n"],
                                     split_type=SplitType.R_PRETRAINED_PROPOSITION.value, min_rl_chunk_size=5,
                                     max_rl_chunk_size=50,enableLLMTouchUp=False)

chunks=chunk_documents_crud_vdb.chunk_documents("extract", file_name=None, file_path="content.txt",splitter_config=splitter_config)

# Step 2: Insert chunks into VDB
chunk_and_upsert_to_vdb("my_index", instructions="split by headings", splitter_config=splitter_config)

# Step 3: Fetch stored chunks
docs = fetch_vdb_chunks_grouped_by_document_name("my_index")

# Step 4: Update chunks if needed
update_vdb_chunks_grouped_by_document_name("my_index", dataset=docs)

---

🛠 Use Cases

• Preprocessing documents for LLM ingestion
• Semantic search and Q&A systems
• Vector database indexing and retrieval
• Maintaining versioned document chunks