feat: Hybrid RAG knowledge base, sentence-aware chunking, MeshMonitor HTTP sync

Knowledge Base:
- Hybrid FTS5 + vector search using sqlite-vec and bge-small-en-v1.5
- Reciprocal Rank Fusion for result merging
- Domain-aware query construction handles typos
- Configurable weights for keyword vs semantic matching

Message Chunking:
- Sentence-aware splitting respects message boundaries
- Continuation prompts for long responses
- Natural follow-up detection (yes, ok, continue, more, etc.)
- Per-user continuation state management

MeshMonitor Integration:
- HTTP API trigger sync (replaces file-based triggers.json)
- Dynamic refresh interval
- Trigger injection into LLM prompt

Other:
- Updated system prompt for better response length control
- Simplified responder to handle message lists
- Updated README with new features and architecture diagram
- Cleaned up config.example.yaml

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
root 2026-05-04 07:44:12 +00:00
commit 0e36869a5f
14 changed files with 986 additions and 464 deletions

View file

@ -1,74 +1,84 @@
# MeshAI Configuration
# LLM-powered Meshtastic assistant
#
# Copy this to config.yaml and customize as needed
# For Docker: mount as /data/config.yaml
# === BOT IDENTITY ===
bot:
name: ai # Bot's display name
owner: "" # Owner's callsign (optional)
respond_to_dms: true # Respond to direct messages
filter_bbs_protocols: true # Ignore advBBS sync/notification messages
# === MESHTASTIC CONNECTION ===
connection:
type: tcp # serial | tcp
serial_port: /dev/ttyUSB0 # For serial connection
tcp_host: localhost # For TCP connection (meshtasticd)
tcp_port: 4403
# === RESPONSE BEHAVIOR ===
response:
delay_min: 2.2 # Min delay before responding (seconds)
delay_max: 3.0 # Max delay before responding
max_length: 150 # Max chars per message chunk
max_messages: 2 # Max message chunks per response
# === CONVERSATION HISTORY ===
history:
database: /data/conversations.db
max_messages_per_user: 50 # Messages to keep per user
conversation_timeout: 86400 # Conversation expiry (seconds, 86400=24h)
auto_cleanup: true # Auto-delete old conversations
cleanup_interval_hours: 24 # How often to run cleanup
max_age_days: 30 # Delete conversations older than this
# === MEMORY OPTIMIZATION ===
memory:
enabled: true # Enable rolling summary memory
window_size: 4 # Recent message pairs to keep in full
summarize_threshold: 8 # Messages before re-summarizing
# === MESH CONTEXT ===
context:
enabled: true # Observe channel traffic for LLM context
observe_channels: [] # Channel indices to observe (empty = all)
ignore_nodes: [] # Node IDs to exclude from observation
max_age: 2592000 # Max age in seconds (default 30 days)
max_context_items: 20 # Max observations injected into LLM context
# === LLM BACKEND ===
llm:
backend: openai # openai | anthropic | google
api_key: "" # API key (or use LLM_API_KEY env var)
base_url: https://api.openai.com/v1 # API base URL
model: gpt-4o-mini # Model name
timeout: 30 # Request timeout (seconds)
system_prompt: >-
You are a helpful assistant on a Meshtastic mesh network.
Keep responses VERY brief - under 250 characters total.
Be concise but friendly. No markdown formatting.
google_grounding: false # Enable Google Search grounding (Gemini only, $35/1k queries)
# === WEATHER ===
weather:
primary: openmeteo # openmeteo | wttr | llm
fallback: llm # openmeteo | wttr | llm | none
default_location: "" # Default location for !weather (optional)
# === MESHMONITOR INTEGRATION ===
meshmonitor:
enabled: false
triggers_file: /data/triggers.json
inject_into_prompt: true
# MeshAI Configuration
# LLM-powered Meshtastic assistant
#
# Copy this to config.yaml and customize as needed
# For Docker: mount as /data/config.yaml
# === BOT IDENTITY ===
bot:
name: ai # Bot's display name
owner: "" # Owner's callsign (optional)
respond_to_dms: true # Respond to direct messages
filter_bbs_protocols: true # Ignore advBBS sync/notification messages
# === MESHTASTIC CONNECTION ===
connection:
type: tcp # serial | tcp
serial_port: /dev/ttyUSB0 # For serial connection
tcp_host: localhost # For TCP connection (meshtasticd)
tcp_port: 4403
# === RESPONSE BEHAVIOR ===
response:
delay_min: 2.2 # Min delay before responding (seconds)
delay_max: 3.0 # Max delay before responding
max_length: 200 # Max chars per message chunk
max_messages: 3 # Max message chunks per response
# === CONVERSATION HISTORY ===
history:
database: /data/conversations.db
max_messages_per_user: 50 # Messages to keep per user
conversation_timeout: 86400 # Conversation expiry (seconds, 86400=24h)
auto_cleanup: true # Auto-delete old conversations
cleanup_interval_hours: 24 # How often to run cleanup
max_age_days: 30 # Delete conversations older than this
# === MEMORY OPTIMIZATION ===
memory:
enabled: true # Enable rolling summary memory
window_size: 4 # Recent message pairs to keep in full
summarize_threshold: 8 # Messages before re-summarizing
# === MESH CONTEXT ===
context:
enabled: true # Observe channel traffic for LLM context
observe_channels: [] # Channel indices to observe (empty = all)
ignore_nodes: [] # Node IDs to exclude from observation
max_age: 2592000 # Max age in seconds (default 30 days)
max_context_items: 20 # Max observations injected into LLM context
# === LLM BACKEND ===
llm:
backend: openai # openai | anthropic | google
api_key: "" # API key (or use LLM_API_KEY env var)
base_url: https://api.openai.com/v1 # API base URL
model: gpt-4o-mini # Model name
timeout: 30 # Request timeout (seconds)
system_prompt: >-
You are a helpful assistant on a Meshtastic mesh network.
Keep responses very brief - 1-2 short sentences, under 300 characters.
Only give longer answers if the user explicitly asks for detail or explanation.
Be concise but friendly. No markdown formatting.
google_grounding: false # Enable Google Search grounding (Gemini only, $35/1k queries)
# === WEATHER ===
weather:
primary: openmeteo # openmeteo | wttr | llm
fallback: llm # openmeteo | wttr | llm | none
default_location: "" # Default location for !weather (optional)
# === MESHMONITOR INTEGRATION ===
meshmonitor:
enabled: false # Enable MeshMonitor trigger sync
url: "" # MeshMonitor web UI URL (e.g. http://192.168.1.100:8080)
inject_into_prompt: true # Include trigger list in LLM prompt
refresh_interval: 300 # Seconds between trigger refreshes
# === KNOWLEDGE BASE (RAG) ===
knowledge:
enabled: false # Enable knowledge base search
db_path: "" # Path to knowledge SQLite database
top_k: 5 # Number of chunks to retrieve per query
fts_weight: 0.5 # Weight for FTS5 keyword matches (0-1)
vector_weight: 0.5 # Weight for vector semantic matches (0-1)