Skip to main content

​
Overview

The EpisodicGistExtraction class performs two-stage extraction to capture both detailed episodic facts and high-level semantic gists. This dual extraction provides both granular information and abstract summaries of episodic content.

​
Class: EpisodicGistExtraction

​
Constructor

from remem.information_extraction.episodic_gist_extraction_openai import EpisodicGistExtraction
from remem.llm.openai_gpt import CacheOpenAI

llm_model = CacheOpenAI(model="gpt-4")
extractor = EpisodicGistExtraction(llm_model=llm_model, global_config=config)
​
llm_model
CacheOpenAI
required
The language model instance used for extraction
​
global_config
object
default:"None"
Optional global configuration containing:
  • dataset (str): Dataset name for template selection
  • seed (int): Random seed for reproducibility
  • temperature (float): LLM temperature parameter

​
Methods

​
batch_openie()

Extract both gists and facts from multiple episodes in a two-stage process. 
def batch_openie(
    self,
    chunks: Dict[str, ChunkInfo]
) -> Tuple[Dict[str, EpisodeRawOutput]]

​
Parameters

​
chunks
Dict[str, ChunkInfo]
required
Dictionary of chunks to process. Each key is a chunk ID, and each value contains:
  • metadata (dict): Episode metadata for passage construction
  • Other ChunkInfo fields (see OpenIE documentation)

​
Returns

A dictionary of EpisodeRawOutput objects:
​
EpisodeRawOutput
object
Combined extraction results containing:
  • chunk_id (str): The chunk identifier
  • verbatim (str): Original passage text
  • facts (List[str]): Detailed episodic facts extracted
  • gists (List[str]): High-level semantic summaries
  • response (str): Raw LLM response
  • metadata (dict): Token usage and processing information

​
Example Usage

from remem.information_extraction.episodic_gist_extraction_openai import EpisodicGistExtraction
from remem.llm.openai_gpt import CacheOpenAI

# Initialize the extractor
llm_model = CacheOpenAI(model="gpt-4")
config = type('Config', (), {
    'dataset': 'locomo',
    'seed': 42,
    'temperature': 0.0
})()

extractor = EpisodicGistExtraction(llm_model=llm_model, global_config=config)

# Prepare episodic chunks
chunks = {
    "episode_1": {
        "metadata": {
            "timestamp": "2024-01-15 14:30",
            "content": "User discussed their travel plans to Japan, mentioning interest in visiting Kyoto temples and trying authentic ramen. They have a budget of $3000 and plan to go in April."
        },
        "num_tokens": 80,
        "content": "User discussed travel plans...",
        "chunk_order": [(0, 1)],
        "full_doc_ids": ["session_456"]
    }
}

# Extract gists and facts
results = extractor.batch_openie(chunks)

# Access results
for chunk_id, episode_output in results.items():
    print(f"\nEpisode: {chunk_id}")
    print(f"Verbatim: {episode_output.verbatim}")
    
    print("\nGists:")
    for gist in episode_output.gists:
        print(f"  - {gist}")
    # Output:
    #   - User is planning a trip to Japan
    #   - User wants cultural and culinary experiences
    #   - User has specific budget and timing constraints
    
    print("\nFacts:")
    for fact in episode_output.facts:
        print(f"  - {fact}")
    # Output:
    #   - User plans to visit Japan
    #   - User wants to visit Kyoto temples
    #   - User wants to try authentic ramen
    #   - User has a budget of $3000
    #   - User plans to travel in April

​
Two-Stage Extraction Process

​
Stage 1: Gist Extraction

  1. Uses episodic_gist_extraction prompt template
  2. Extracts high-level semantic summaries
  3. Captures abstract themes and main points
  4. Processes all chunks in parallel

​
Stage 2: Fact Extraction

  1. Uses episodic_fact_extraction prompt template
  2. Includes previously extracted gists as context
  3. Extracts detailed, specific facts
  4. Leverages gists to inform fact granularity

​
Context Enhancement

When extracting facts, the previously extracted gists are included in the prompt: 
Previously extracted gists for this session:
- [Gist 1]
- [Gist 2]
...
This helps the model:
  • Maintain consistency between abstractions and details
  • Extract facts at an appropriate level of granularity
  • Avoid redundancy between gists and facts

​
Template Selection

Automatic template selection based on dataset:
  • Wikipedia-based datasets: Uses *_wikipedia templates
    • menatqa, timeqa, musique, complex_tr, 2wikimultihopqa
  • Conversation datasets: Uses *_locomo templates (default)
  • Custom datasets: Matches by dataset prefix

​
Output Structure

​
Gists

High-level semantic summaries that:
  • Capture the main themes or topics
  • Abstract away specific details
  • Provide context for understanding the episode
  • Are typically 1-3 sentences each

​
Facts

Detailed, specific information that:
  • Reference concrete entities, dates, numbers
  • Capture precise relationships and actions
  • Can be verified against the verbatim text
  • Are more granular than gists

​
Performance Considerations

  • Two-stage processing: Gists extracted first, then facts
  • Parallel execution: Each stage processes all chunks concurrently
  • Configurable workers: Default max_workers=10 for ThreadPoolExecutor
  • Ordered results: Maintains original chunk ordering
  • Progress tracking: Separate progress bars for gists and facts
  • Token metrics: Tracks usage for both extraction stages

​
Error Handling

  • Graceful handling of JSON parsing errors
  • Automatic fixing of truncated JSON responses
  • Null value replacement with empty strings
  • Error metadata attached to results
  • Logging of extraction exceptions
  • Returns empty lists on failure while preserving chunk structure

​
Use Cases

  • Memory systems: Store both detailed and summarized episode representations
  • Question answering: Retrieve relevant gists first, then drill down to facts
  • Conversation analysis: Understand both high-level topics and specific details
  • Long-term memory: Compress episodic information at multiple granularities
  • Hierarchical retrieval: Enable multi-level semantic search

Build docs developers (and LLMs) love

Get started for freeTalk to us