Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/intuit-ai-research/REMem/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

Indexing is the first step in using REMem. The index() method processes your documents, extracts structured information (entities, facts, and relationships), creates embeddings, and builds a knowledge graph for efficient retrieval.

​
Basic Usage

1

Initialize ReMem

Create a ReMem instance with your configuration:
from remem import ReMem
from remem.utils.config_utils import BaseConfig

config = BaseConfig(
    llm_name="gpt-4o-mini",
    embedding_model_name="nvidia/NV-Embed-v2",
    dataset="my_dataset"
)

remem = ReMem(global_config=config, working_dir="./remem_data")
2

Prepare Documents

Format your documents as a list of strings:
docs = [
    "Machine learning is a subset of artificial intelligence.",
    "Neural networks are inspired by biological neurons.",
    "Deep learning uses multiple layers of neural networks."
]
3

Index Documents

Call the index() method to process and store your documents:
remem.index(docs)
This performs:
  • Document chunking (if needed)
  • Information extraction (entities, facts)
  • Embedding generation
  • Knowledge graph construction

​
Complete Example from main.py

Here’s a real-world example from the REMem codebase:
main.py
import json
from remem import ReMem
from remem.utils.config_utils import BaseConfig

# Load corpus
corpus_path = f"reproduce/dataset/musique_corpus.json"
with open(corpus_path, "r") as f:
    corpus = json.load(f)

# Format documents
docs = [f"{doc['title']}\n{doc['text']}" for doc in corpus]

# Configure ReMem
config = BaseConfig(
    llm_name="gpt-4o-mini",
    embedding_model_name="nvidia/NV-Embed-v2",
    dataset="musique",
    force_index_from_scratch=False,  # Reuse existing index if available
    retrieval_top_k=200,
    qa_top_k=5
)

remem = ReMem(global_config=config, working_dir="./outputs/musique")

# Index documents
remem.index(docs)

​
Extraction Methods

REMem supports multiple information extraction strategies via extract_method:

​
OpenIE (Default)

Extracts subject-predicate-object triples: 
config = BaseConfig(
    extract_method="openie",  # Default
    llm_name="gpt-4o-mini"
)

​
Episodic

Optimized for conversational/temporal data: 
config = BaseConfig(
    extract_method="episodic",
    llm_name="gpt-4o-mini"
)

​
Episodic Gist

Extracts gists, facts, entities, and metadata: 
config = BaseConfig(
    extract_method="episodic_gist",
    llm_name="gpt-4o-mini"
)

​
Temporal

Captures time-sensitive relationships: 
config = BaseConfig(
    extract_method="temporal",
    llm_name="gpt-4o-mini"
)

​
Incremental Indexing

By default, REMem supports incremental indexing. New documents are added to the existing graph: 
# First batch
docs_batch_1 = ["Document 1", "Document 2"]
remem.index(docs_batch_1)

# Second batch (incremental)
docs_batch_2 = ["Document 3", "Document 4"]
remem.index(docs_batch_2)  # Adds to existing graph
Set force_index_from_scratch=True in BaseConfig to rebuild the entire index from scratch. This will ignore any existing embeddings and graph data.

​
Indexing Performance

​
Online vs Offline Mode

Control LLM inference mode with llm_infer_mode: 
# Online mode (default) - real-time API calls
config = BaseConfig(
    llm_infer_mode="online",
    llm_name="gpt-4o-mini"
)

# Offline mode - batch processing with vLLM
config = BaseConfig(
    llm_infer_mode="offline",
    llm_name="meta-llama/Llama-3.3-70B-Instruct",
    vllm_tensor_parallel_size=2
)

​
Batch Sizes

Adjust embedding batch size for optimal performance: 
config = BaseConfig(
    embedding_batch_size=16,  # Default: 16
    embedding_model_name="nvidia/NV-Embed-v2"
)

​
Chunking Configuration

Control document preprocessing with chunking parameters: 
config = BaseConfig(
    preprocess_chunk_max_token_size=512,  # Max tokens per chunk
    preprocess_chunk_overlap_token_size=128,  # Overlap between chunks
    preprocess_encoder_name="gpt-4o"  # Tokenizer for chunking
)
If preprocess_chunk_max_token_size=None, documents are treated as single chunks without splitting.

​
Graph Construction Parameters

Customize graph building behavior: 
config = BaseConfig(
    # Synonymy edge construction
    synonymy_edge_topk=2047,  # Number of neighbors for synonym detection
    synonymy_edge_sim_threshold=0.8,  # Similarity threshold for synonyms
    
    # Graph type
    is_directed_graph=False,  # Use undirected graph (default)
    graph_type="facts_and_sim_passage_node_unidirectional"
)

​
Saving OpenIE Results

Persist extracted information for analysis or reuse: 
config = BaseConfig(
    save_openie=True,  # Default: saves extraction results
    force_openie_from_scratch=False  # Reuse existing OpenIE results
)

​
What Gets Created

After indexing, REMem creates the following in your working_dir:
  • chunk_embeddings/ - Embeddings for document chunks
  • entity_embeddings/ - Embeddings for extracted entities
  • fact_embeddings/ - Embeddings for extracted facts
  • graph.pkl - Knowledge graph structure
  • openie_results_*.json - Extracted entities and relationships (if save_openie=True)

​
Next Steps

Retrieval

Learn how to retrieve relevant passages

Configuration

Explore all configuration options

Build docs developers (and LLMs) love

Get started for freeTalk to us