config.yaml

# =========================
# CLEVERBEE CONFIGURATION
# =========================
# This file controls the core behavior, model selection, data processing,
# tool usage, and operational parameters of the CleverBee system.

# ===================================
# SECTION 1: CORE MODEL CONFIGURATION
# ===================================
# Settings determining the primary AI models used for reasoning, planning,
# and final report generation.

# --- Primary Model Type ---
# Selects which type of model to use as the primary reasoning agent.
# This setting controls which model handles research planning, tool selection, and report writing.
# Options:
# - "gemini": Use Google's Gemini models via API
# - "claude": Use Anthropic's Claude models via API
# - "local": Use a locally hosted GGUF model via llama.cpp
PRIMARY_MODEL_TYPE: "gemini"
# --- Specific Model Selection ---
# Define the exact model name to use for each provider type.
# The system will use the model that corresponds to the selected PRIMARY_MODEL_TYPE.
CLAUDE_MODEL_NAME: "claude-3-7-sonnet-20250219"
GEMINI_MODEL_NAME: "gemini-2.5-pro-preview-03-25"
LOCAL_MODEL_NAME: null
LOCAL_MODEL_QUANT_LEVEL: null
# --- Local Model Usage ---
# Determines if locally hosted models should be prioritized for summarization tasks.
# This setting is automatically updated by `setup.sh` based on the chosen summarizer model.
# If `true`, the system attempts to use the locally downloaded model specified below.
# If `false`, the cloud model (e.g., `gemini-2.0-flash`) is used for summarization.
USE_LOCAL_SUMMARIZER_MODEL: false
# --- Local Model Configuration ---
# Directory where local model files (like GGUF) are stored.
# This path is used by the LLM factory when USE_LOCAL_SUMMARIZER_MODEL is true.
LOCAL_MODELS_DIR: "models"
# --- GPU Offloading (for Local Models) ---
# Number of model layers to offload to the GPU (e.g., Metal on Mac, CUDA on Nvidia).
# - `0` uses CPU only.
# - `-1` attempts to offload all layers (may use excessive RAM or be unstable).
# - A positive number (e.g., 1, 10, 32) offloads that many layers.
# `setup.sh` attempts to detect your hardware (Apple Silicon, Nvidia) and
# recommends an optimal value during setup, which you can accept or override.
# Recommended starting points (if setting manually):
#   - Apple Silicon (M1/M2/M3): 10-48 depending on model & chip (see setup.sh logic)
#   - Nvidia GPU: 20-35 depending on model & VRAM
N_GPU_LAYERS: 0
# =======================================
# SECTION 2: SUMMARIZATION CONFIGURATION
# =======================================
# Settings controlling how text content (e.g., web pages, documents) is summarized
# during the research process, both for intermediate steps and potential memory management.

# --- Summarization Model ---
# Specifies which model to use for summarization tasks.
# This value is set by `setup.sh` based on your selection.
# Available options presented during setup depend on detected RAM and GPU:
#   - gemini-2.0-flash:                          [Cloud, Recommended for <8GB RAM, no download]
#   - DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf: [Local GGUF, Recommended for NVIDIA and >=8GB RAM]
#   - yarn-mistral-7b-64k.Q4_K_M.gguf:        [Local GGUF, Optional for NVIDIA users >=8GB RAM]
# Selecting a local model sets `USE_LOCAL_SUMMARIZER_MODEL` to `true` automatically by setup.sh.
SUMMARIZER_MODEL: "gemini-2.0-flash"
# --- General Summarization Limits ---
# Sets the maximum number of tokens the summarization model should generate for EACH summary.
# This applies to both cloud and local summarization models. It directly influences
# the length and detail of intermediate summaries produced by the ContentManager.
SUMMARY_MAX_TOKENS: 1200
# --- Final Summary Limits ---
# Sets the maximum number of tokens the final summarization model should generate for the FINAL report.
# This applies specifically to the LLM instance responsible for creating the comprehensive research report.
# This is defined in the ResearcherAgent's initialization.
FINAL_SUMMARY_MAX_TOKENS: 16000
# --- Post-Research Conversation Limits ---
# Sets the maximum tokens for LLM responses during follow-up after research is complete,
# especially when requesting the full accumulated content.
# Default: 32768
MAX_CONTINUED_CONVERSATION_TOKENS: 32768
# --- Next Step Model ---
# Specifies which model to use for next step/agentic decision tasks.
# This model is used for agentic decision steps (the action loop), separate from the primary LLM and summarizer model.
# Example: "gemini-2.5-flash" (recommended for fast, cost-effective next-step reasoning)
NEXT_STEP_MODEL: "gemini-2.5-flash-preview-04-17"
# If true, next step model will default to "thinking mode" (extra tokens/depth) unless overridden by planner output.
NEXT_STEP_THINKING_DEFAULT: false
# ======================================
# SECTION 3: CONTENT PROCESSING & MEMORY
# ======================================
# Settings related to how raw text content is handled, chunked, loaded,
# and how the agent's conversation history is managed.

# --- Text Chunking Parameters ---
# When processing large documents with local models, the text is split into smaller chunks.
# These settings, used by the ContentManager's RecursiveCharacterTextSplitter, control this process.
# The `CHUNK_SIZE` is adjusted automatically by `setup.sh` based on the selected local model:
#   - gemini-2.0-flash:                          CHUNK_SIZE = 0 (no chunking, cloud model)
#   - DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf: CHUNK_SIZE = 3000
#   - yarn-mistral-7b-64k.Q4_K_M.gguf:         CHUNK_SIZE = 7000
CHUNK_SIZE: 0
CHUNK_OVERLAP: 400 # The number of characters from the end of one chunk to include at the start of the next.
# --- Content Optimization ---
# Options to potentially improve performance or change how content is presented to the LLM.
USE_PROGRESSIVE_LOADING: true # If true, the ContentManager might initially provide summaries
# instead of full content, loading full content only when needed.

# --- Agent Reasoning Enhancement ---
# Settings potentially allowing the agent more 'thinking time' or resources for complex steps.
# NOTE: The 'ENABLE_THINKING' feature's effectiveness might depend on the specific LLM used
# (e.g., Claude 3.7 was mentioned in code comments).
ENABLE_THINKING: true # If true, attempts to allocate more processing capacity (e.g., tokens)
# for specific reasoning steps, potentially improving complex task performance.
THINKING_BUDGET: 4000 # Maximum *additional* tokens to allocate for a thinking step if ENABLE_THINKING is true.
# --- Conversation Memory ---
# Configures the agent's memory, managed by ConversationSummaryBufferMemory.
# This memory stores recent interactions directly and summarizes older ones.
# CONVERSATION_MEMORY_MAX_TOKENS: 120000  # Example: For Claude 3.7 Sonnet
CONVERSATION_MEMORY_MAX_TOKENS: 900000 # Example: For Gemini 2.5 Pro
# Set to ~90% of the main reasoning model's context window to allow space for the prompt,
# tool usage details, and the model's response generation.
# Adjust based on the MAIN_LLM_PROVIDER and selected model's context size.

# ======================================
# SECTION 4: SEARCH & BROWSER SETTINGS
# ======================================
# Parameters controlling how the agent gathers information from the web and specific platforms.

# --- Web Search Limits ---
# Controls the number of standard web search results (e.g., from Google) processed per research run.
# This applies to the general `web_browser` tool when used for searching or extracting from search results.
MIN_REGULAR_WEB_PAGES: 1 # The minimum number of distinct web pages the agent should aim to process.
MAX_REGULAR_WEB_PAGES: 10 # The maximum number of distinct web pages the agent is allowed to process.
# This acts as a cap to control scope and cost.

# --- Reddit Search Limits ---
# Controls the number of Reddit posts retrieved when using the `reddit_search` tool.
MIN_POSTS_PER_SEARCH: 1 # Minimum Reddit posts the tool should attempt to return per search query.
MAX_POSTS_PER_SEARCH: 10 # Maximum Reddit posts the tool should return per search query.
# --- Browser Behavior ---
# Settings for the underlying Playwright browser instance used by tools like `web_browser`.
BROWSER_NAVIGATION_TIMEOUT: 15 # Maximum time (in seconds) to wait for a single page
# navigation (e.g., clicking a link, loading a URL) to complete.
USE_CAPTCHA_SOLVER: true # If true, attempts to use an integrated CAPTCHA solving service
# (like Recognizer, if available) when encountering challenges.
CAPTCHA_SOLVER_TIMEOUT: 2000 # Maximum time (in milliseconds) to wait for the CAPTCHA solver
# to attempt a solution before giving up on that attempt.

# ===================================
# SECTION 5: TOOL CONFIGURATION
# ===================================
# Enables or disables specific tools available to the agent.
# The agent's LLM decides which enabled tools to use based on the task.
# Disabling a tool here prevents the agent from using it.
# This corresponds to the TOOLS_CONFIG variable used in tool loading.
tools:
  web_browser:
    enabled: true # Enables the core web browsing/extraction tool (Playwright based).
    # Used for general web searches and fetching content from URLs.
  reddit_search:
    enabled: true # Enables the tool specifically designed for searching Reddit.
  reddit_extract_post:
    enabled: true # Enables the tool for extracting content from a specific Reddit post URL.
# ==========================================
# SECTION 6: USAGE TRACKING & PRICING
# ==========================================
# Settings related to monitoring resource consumption (LLM tokens) and estimating costs.

# --- Tracking ---
TRACK_TOKEN_USAGE: true # If true, enables the TokenUsageCallbackHandler to count input/output
# tokens for each LLM call made through LangChain.
LOG_COST_SUMMARY: true # If true (and TRACK_TOKEN_USAGE is true), prints a summary of
# estimated costs based on token counts and configured prices
# at the end of each research session.

# --- Pricing (per 1000 tokens) ---
# These values are used ONLY IF TRACK_TOKEN_USAGE is true.
# They define the estimated cost for input and output tokens for different models.
# Ensure these reflect the actual pricing of the models you are using.

# Gemini 2.5 Pro Pricing
GEMINI_COST_PER_1K_INPUT_TOKENS: 0.00125
GEMINI_COST_PER_1K_OUTPUT_TOKENS: 0.01000
# Gemini 2.0 Flash Pricing
GEMINI_FLASH_COST_PER_1K_INPUT: 0.00010
GEMINI_FLASH_COST_PER_1K_OUTPUT: 0.00040
# Claude 3.7 Sonnet Pricing
CLAUDE_COST_PER_1K_INPUT_TOKENS: 0.008
CLAUDE_COST_PER_1K_OUTPUT_TOKENS: 0.024
# ======================================
# SECTION 7: LOGGING CONFIGURATION
# ======================================
# Controls the verbosity of application logs.

# --- Log Level ---
# Sets the minimum severity level for log messages that will be outputted.
# Affects all loggers in the application unless they have a specific level set.
# Options (most to least verbose): "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"
LOG_LEVEL: "DEBUG"