Overview The BaseConfig class in src/remem/utils/config_utils.py contains all configuration parameters for REMem. This guide documents every parameter with examples and recommended values.
Basic Setup from remem.utils.config_utils import BaseConfig config = BaseConfig( llm_name = "gpt-4o-mini" , embedding_model_name = "nvidia/NV-Embed-v2" , dataset = "my_dataset" )
LLM Parameters Model Selection config = BaseConfig( # Primary LLM for extraction and QA llm_name = "gpt-4o-mini" , # Default: "gpt-4o-mini" # Separate LLMs for specific tasks extract_llm_label = None , # Defaults to llm_name qa_llm_label = None , # Defaults to llm_name # LLM API endpoint llm_base_url = None , # Default: None (uses OpenAI) # Azure OpenAI use_azure = False # Set to True for Azure )
Generation Parameters config = BaseConfig( # Token limits max_new_tokens = 2048 , # Default: 2048, None = model default max_model_len = 4096 , # Context window for vLLM # Sampling temperature = 0 , # Default: 0 (deterministic) seed = None , # Random seed for reproducibility num_gen_choices = 1 , # Completions per prompt # Response format extract_format = None , # "json_object" or "json_schema" )
For Llama models, set extract_format="json_schema" to ensure proper structured output.
Inference Mode config = BaseConfig( # Online: real-time API calls llm_infer_mode = "online" , # Default # Offline: batch processing with vLLM # llm_infer_mode="offline", # vllm_tensor_parallel_size=2, # GPU parallelism # max_num_seqs=256 # Max concurrent sequences )
Async Configuration config = BaseConfig( max_retries = 10 , # Retry attempts for async API calls )
Embedding Parameters Model Selection config = BaseConfig( embedding_model_name = "nvidia/NV-Embed-v2" , # Default # Other options: # "text-embedding-3-large" # OpenAI # "text-embedding-3-small" # OpenAI # "GritLM/GritLM-7B" # GritLM )
Embedding Configuration config = BaseConfig( embedding_batch_size = 16 , # Batch size for encoding embedding_return_as_normalized = True , # Normalize embeddings embedding_max_seq_len = 2048 , # Max sequence length )
Storage Parameters Index Management config = BaseConfig( # Rebuild index from scratch force_index_from_scratch = False , # Default: reuse existing # Rebuild OpenIE results force_openie_from_scratch = False , # Default: reuse existing # Save extracted information save_openie = True , # Default: saves to disk )
Save Directory config = BaseConfig( save_dir = None , # Auto-generated if None dataset = "musique" , # Used to create save_dir ) # Auto-generated paths: # - dataset="musique" → "outputs/musique" # - dataset="locomo_v1" → "outputs/locomo/locomo_v1" # - dataset=None → "outputs"
config = BaseConfig( extract_method = "openie" , # Default: OpenIE triples # Options: "openie", "episodic", "episodic_gist", "temporal" )
Method Descriptions:
openie - Extracts subject-predicate-object triplesepisodic - Conversational/temporal data extractionepisodic_gist - Gists, facts, entities, and metadatatemporal - Time-sensitive relationship extraction
Skip Graph Construction config = BaseConfig( skip_graph = False , # Set True to only do OpenIE )
Set skip_graph=True when running vLLM offline indexing for the first time to extract information without building the graph.
Text Preprocessing Chunking Parameters config = BaseConfig( text_preprocessor_class_name = "TextPreprocessor" , # Preprocessor class # Tokenizer for chunking preprocess_encoder_name = "gpt-4o" , # Chunk size preprocess_chunk_max_token_size = None , # None = no chunking # Overlap between chunks preprocess_chunk_overlap_token_size = 128 , # Chunking function preprocess_chunk_func = "by_token" , )
Example: Enable Chunking config = BaseConfig( preprocess_chunk_max_token_size = 512 , # 512 tokens per chunk preprocess_chunk_overlap_token_size = 128 , # 128 token overlap preprocess_encoder_name = "gpt-4o" # Use GPT-4o tokenizer )
Graph Construction Graph Type config = BaseConfig( graph_type = "facts_and_sim_passage_node_unidirectional" , # Default # Options: # - "dpr_only" # - "facts_and_sim" # - "facts_and_sim_passage_node_unidirectional" is_directed_graph = False , # Use undirected graph )
Synonymy Edges Control entity linking via similarity:
config = BaseConfig( # KNN parameters for synonym detection synonymy_edge_topk = 2047 , # Neighbors to consider synonymy_edge_query_batch_size = 1000 , # Query batch size synonymy_edge_key_batch_size = 10000 , # Key batch size # Similarity threshold synonymy_edge_sim_threshold = 0.8 , # 0.0-1.0 )
Episodic Graph Options config = BaseConfig( # Concatenate gists per chunk into single node concatenate_gists_per_chunk = False , # Split long verbatim content into separate nodes split_verbatim_per_chunk = True , )
Retrieval Parameters Basic Retrieval config = BaseConfig( retrieval_top_k = 200 , # Total passages to retrieve linking_top_k = 5 , # Entities to link per step damping = 0.5 , # PPR damping factor (0-1) )
Passage Node Weight Balance passage vs entity importance:
config = BaseConfig( passage_node_weight = 0.05 , # Multiplicative factor for passages in PPR )
Reranking config = BaseConfig( rerank_dspy_file_path = None , # Path to trained DSPy filter # Example: # rerank_dspy_file_path="src/remem/prompts/dspy_prompts/filter_llama3.3-70B-Instruct.json" )
QA Parameters Reading Configuration config = BaseConfig( qa_top_k = 5 , # Passages to feed to LLM qa_passage_prefix = "Wikipedia Title: " , # Prefix for each passage qa_prompt_template = None , # Custom prompt template name qa_reader = "remem" , # "remem" or "tiser" )
QA Prompt Template # Auto-detect based on dataset config = BaseConfig( dataset = "musique" , qa_prompt_template = None # Uses "rag_qa_musique" template ) # Or specify explicitly config = BaseConfig( qa_prompt_template = "rag_qa_custom" )
Agent Parameters For episodic and temporal extraction methods:
config = BaseConfig( # Tool selection mode agent_fixed_tools = False , # False = full tool selection # Reasoning steps agent_max_steps = 5 , # Max steps for agent reasoning # Fixed tool mode configuration agent_fixed_retrieval_tool = "semantic_retrieve" , # or "lexical_retrieve" )
Fixed Tools Mode:
agent_max_steps=1 - Semantic retrieve onlyagent_max_steps=2 - Semantic retrieve + output answer
Evaluation Parameters Enable Evaluation config = BaseConfig( do_eval_retrieval = True , # Evaluate retrieval quality do_eval_qa = True , # Evaluate answer quality )
Corpus Configuration config = BaseConfig( corpus_len = None , # Total number of documents in corpus )
Complete Configuration Examples Minimal Configuration config = BaseConfig( llm_name = "gpt-4o-mini" , embedding_model_name = "nvidia/NV-Embed-v2" )
Production Configuration config = BaseConfig( # LLM llm_name = "gpt-4o-mini" , extract_llm_label = "gpt-4o-mini" , qa_llm_label = "gpt-4o" , # More powerful for QA max_new_tokens = 2048 , temperature = 0 , # Embeddings embedding_model_name = "nvidia/NV-Embed-v2" , embedding_batch_size = 16 , # Storage dataset = "my_dataset" , force_index_from_scratch = False , save_openie = True , # Preprocessing preprocess_chunk_max_token_size = 512 , preprocess_chunk_overlap_token_size = 128 , # Graph graph_type = "facts_and_sim_passage_node_unidirectional" , synonymy_edge_topk = 2047 , synonymy_edge_sim_threshold = 0.8 , # Retrieval retrieval_top_k = 200 , linking_top_k = 5 , damping = 0.5 , passage_node_weight = 0.05 , rerank_dspy_file_path = "path/to/filter.json" , # QA qa_top_k = 5 , qa_passage_prefix = "Wikipedia Title: " , # Evaluation do_eval_retrieval = True , do_eval_qa = True , )
Offline (vLLM) Configuration config = BaseConfig( # Offline inference llm_infer_mode = "offline" , llm_name = "meta-llama/Llama-3.3-70B-Instruct" , extract_format = "json_schema" , vllm_tensor_parallel_size = 2 , max_num_seqs = 256 , max_model_len = 4096 , # First pass: extract only skip_graph = True , # Other settings... embedding_model_name = "nvidia/NV-Embed-v2" , dataset = "my_dataset" , )
Episodic Gist Configuration config = BaseConfig( # Extraction extract_method = "episodic_gist" , # Agent agent_fixed_tools = False , agent_max_steps = 5 , # Graph options concatenate_gists_per_chunk = False , split_verbatim_per_chunk = True , # Other settings... llm_name = "gpt-4o-mini" , embedding_model_name = "nvidia/NV-Embed-v2" , retrieval_top_k = 200 , qa_top_k = 5 , )
Next Steps
Indexing Learn how to index documents
Embeddings Configure embedding models