Class: SwarmSDK::ContextCompactor

Inherits:
Object
  • Object
show all
Defined in:
lib/swarm_sdk/context_compactor.rb,
lib/swarm_sdk/context_compactor/metrics.rb,
lib/swarm_sdk/context_compactor/token_counter.rb

Overview

ContextCompactor implements intelligent conversation history compression

The Hybrid Production Strategy combines three compression techniques:

  1. Tool result pruning - Aggressively truncate tool outputs (80% of tokens!)
  2. Checkpoint creation - LLM-generated summaries of conversation chunks
  3. Sliding window - Keep recent messages in full detail

Usage

# From Agent::Chat
metrics = chat.compact_context

# With options
metrics = chat.compact_context(
tool_result_max_length: 500,
checkpoint_threshold: 50,
sliding_window_size: 20,
summarization_model: "claude-3-haiku-20240307"
)

Metrics

Returns a Metrics object with compression stats:

  • original_message_count / compressed_message_count
  • original_tokens / compressed_tokens
  • compression_ratio (e.g., 0.15 = 15% of original)
  • messages_removed / messages_summarized
  • time_taken

Defined Under Namespace

Classes: Metrics, TokenCounter

Constant Summary collapse

DEFAULT_OPTIONS =

Default configuration

{
  tool_result_max_length: 500, # Truncate tool results to N chars
  checkpoint_threshold: 50,        # Create checkpoint after N messages
  sliding_window_size: 20,         # Keep last N messages in full
  summarization_model: "claude-3-haiku-20240307", # Fast model for summaries
  preserve_system_messages: true,  # Always keep system messages
  preserve_error_messages: true,   # Always keep error messages
}.freeze

Instance Method Summary collapse

Constructor Details

#initialize(chat, options = {}) ⇒ ContextCompactor

Initialize compactor for a chat instance

Parameters:

  • chat (Agent::Chat) —

    The chat instance to compact

  • options (Hash) (defaults to: {}) —

    Configuration options (see DEFAULT_OPTIONS)



48
49
50
51
52
 
# File 'lib/swarm_sdk/context_compactor.rb', line 48

def initialize(chat, options = {})
  @chat = chat
  @options = DEFAULT_OPTIONS.merge(options)
  @agent_name = chat.provider.respond_to?(:agent_name) ? chat.provider.agent_name : :unknown
end

Instance Method Details

#compact ⇒ ContextCompactor::Metrics

Compact the conversation history using hybrid production strategy

Returns metrics about the compression operation.

Returns:



59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
 
# File 'lib/swarm_sdk/context_compactor.rb', line 59

def compact
  start_time = Time.now
  original_messages = @chat.messages

  # Emit compression_started event
  LogStream.emit(
    type: "compression_started",
    agent: @agent_name,
    message_count: original_messages.size,
    estimated_tokens: TokenCounter.estimate_messages(original_messages),
  )

  # Step 1: Prune tool results
  pruned = prune_tool_results(original_messages)

  # Step 2: Create checkpoint if needed
  checkpointed = create_checkpoint_if_needed(pruned)

  # Step 3: Apply sliding window
  final_messages = apply_sliding_window(checkpointed)

  # Replace messages in chat
  replace_messages(final_messages)

  # Calculate metrics
  time_taken = Time.now - start_time
  metrics = ContextCompactor::Metrics.new(
    original_messages: original_messages,
    compressed_messages: final_messages,
    time_taken: time_taken,
  )

  # Emit compression_completed event
  LogStream.emit(
    type: "compression_completed",
    agent: @agent_name,
    original_message_count: metrics.original_message_count,
    compressed_message_count: metrics.compressed_message_count,
    original_tokens: metrics.original_tokens,
    compressed_tokens: metrics.compressed_tokens,
    compression_ratio: metrics.compression_ratio,
    messages_removed: metrics.messages_removed,
    messages_summarized: metrics.messages_summarized,
    time_taken: metrics.time_taken.round(3),
  )

  metrics
end