jbergner
440423c5b6
All checks were successful
release-tag / release-image (push) Successful in 2m43s
279 lines
12 KiB
Plaintext
279 lines
12 KiB
Plaintext
# Runtime role: brain = full knowledge system; agent = lightweight source poller using the same binary/image.
|
|
BRAIN_MODE=brain
|
|
|
|
# HTTP
|
|
# Direct binary listen address. Docker Compose listens internally on :8090 and
|
|
# exposes BRAIN_PORT on the host.
|
|
BRAIN_LISTEN_ADDR=:8090
|
|
# Docker Compose host port (not consumed by the binary itself).
|
|
BRAIN_PORT=8090
|
|
# URL that remote/source-agent containers can actually reach. Do not use
|
|
# localhost/127.0.0.1 here for a separate container.
|
|
BRAIN_PUBLIC_URL=
|
|
BRAIN_DATA_DIR=./data
|
|
BRAIN_API_KEY=
|
|
|
|
# Read-only production knowledge; comma-separated paths are supported.
|
|
BRAIN_KNOWLEDGE_DIRS=../glpi-ai-agent/knowledge
|
|
|
|
# Writable staging used only for AI-THINK drafts. The Agent does not mount this path.
|
|
BRAIN_STAGING_DIRS=../glpi-ai-knowledgebase/staging
|
|
|
|
# Optional read-only audit streams. Multiple files may be comma-separated.
|
|
BRAIN_AGENT_RUNS_FILES=../glpi-ai-agent/data/runs.jsonl
|
|
|
|
# Batched disk persistence. AI-THINK files and caches are written first;
|
|
# changed graph rows follow in one SQLite/WAL transaction. A final flush is
|
|
# attempted during shutdown.
|
|
BRAIN_PERSIST_INTERVAL=5m
|
|
|
|
# Ollama pool. OLLAMA_URLS takes precedence over legacy OLLAMA_URL.
|
|
OLLAMA_URL=http://localhost:11434
|
|
OLLAMA_URLS=http://localhost:11434
|
|
OLLAMA_NODE_NAMES=local-gpu
|
|
OLLAMA_NODE_WEIGHTS=1
|
|
OLLAMA_ROUTING_MODE=least_inflight
|
|
OLLAMA_NODE_MAX_INFLIGHT=1
|
|
OLLAMA_NODE_HEALTH_INTERVAL=15s
|
|
OLLAMA_NODE_FAILURE_COOLDOWN=30s
|
|
OLLAMA_NODE_REQUEST_TIMEOUT=8m
|
|
OLLAMA_FAILOVER_ENABLED=true
|
|
# 0 = try all configured nodes.
|
|
OLLAMA_FAILOVER_ATTEMPTS=0
|
|
OLLAMA_REQUIRE_SAME_MODEL_DIGEST=true
|
|
OLLAMA_REQUIRE_EMBEDDING_MODEL=true
|
|
OLLAMA_CHAT_MODEL=qwen3:8b
|
|
OLLAMA_EMBEDDING_MODEL=embeddinggemma
|
|
|
|
# Optional read-only GLPI Knowledge Base ingest.
|
|
GLPI_KB_ENABLED=false
|
|
GLPI_URL=https://glpi.example.invalid
|
|
GLPI_API_VERSION=v2.3
|
|
GLPI_CLIENT_ID=
|
|
GLPI_CLIENT_SECRET=
|
|
GLPI_USERNAME=
|
|
GLPI_PASSWORD=
|
|
GLPI_ALLOW_INSECURE_HTTP=false
|
|
GLPI_TIMEOUT=20s
|
|
GLPI_KB_PATH=auto
|
|
GLPI_KB_FILTER=
|
|
GLPI_KB_LIMIT=500
|
|
GLPI_KB_SYNC_INTERVAL=10m
|
|
GLPI_KB_SOURCE=GLPI Knowledge Base
|
|
|
|
# Runtime controls. Boolean/view/performance values are WebUI defaults.
|
|
# Source lists are initial WebUI defaults only; runtime-settings.json takes precedence afterwards.
|
|
# Empty lists mean unrestricted. Values must exactly match the KB JSON source field.
|
|
BRAIN_LEARNING_ENABLED=true
|
|
BRAIN_THINKING_ENABLED=true
|
|
# Initial exact source selections; empty = all sources.
|
|
BRAIN_LEARNING_SOURCES=
|
|
BRAIN_DISPLAY_SOURCES=
|
|
BRAIN_THINKING_SOURCES=
|
|
BRAIN_DEFAULT_VIEW=neural # neural | honeycomb | constellation
|
|
# 0 = unbegrenzt; Webinterface kann den Wert zur Laufzeit ändern
|
|
BRAIN_MAX_DISPLAY_NODES=0
|
|
# Browser-Rendering auf 30 FPS, geringere DPR und reduzierte Effektkosten optimieren.
|
|
BRAIN_LOW_POWER_MODE=false
|
|
|
|
# Sequential enrichment
|
|
BRAIN_AUTO_ENRICH=true
|
|
BRAIN_SCAN_INTERVAL=5m
|
|
# Safety verification: periodically parse/hash full knowledge content even when path/size/mtime are unchanged. 0 disables.
|
|
BRAIN_KNOWLEDGE_FULL_VERIFY_INTERVAL=6h
|
|
BRAIN_ENRICH_INTERVAL=90s
|
|
BRAIN_ENRICH_BATCH_SIZE=3
|
|
BRAIN_ENRICH_STEP_DELAY=3s
|
|
BRAIN_ENRICH_ANCHORS=48
|
|
# precise = vollständige Cosine-Suche; clustered = Semantic Hashing + exakte Top-K-Prüfung
|
|
BRAIN_PROCESSING_MODE=precise
|
|
BRAIN_CLUSTER_HASH_BITS=24
|
|
BRAIN_CLUSTER_HASH_TABLES=2
|
|
|
|
# Experimental mathematical Knowledge<->Knowledge layer. Reuses embeddings that
|
|
# already exist; edge scoring/layout perform no additional model call.
|
|
BRAIN_VECTOR_GRAPH_ENABLED=false
|
|
BRAIN_VECTOR_GRAPH_NEIGHBORS=4
|
|
BRAIN_VECTOR_GRAPH_CANDIDATES=96
|
|
BRAIN_VECTOR_GRAPH_MIN_SIMILARITY=0.80
|
|
BRAIN_VECTOR_GRAPH_MIN_AFFINITY=0.35
|
|
# Optional second pass for nodes left isolated after the conservative mutual-kNN pass.
|
|
BRAIN_VECTOR_GRAPH_ORPHAN_PASS=false
|
|
BRAIN_VECTOR_GRAPH_ORPHAN_NEIGHBORS=2
|
|
BRAIN_VECTOR_GRAPH_ORPHAN_CANDIDATES=256
|
|
BRAIN_VECTOR_GRAPH_ORPHAN_MIN_SIMILARITY=0.80
|
|
BRAIN_VECTOR_GRAPH_ORPHAN_MIN_AFFINITY=0.30
|
|
# Let an integrated BRAIN_MODE=agent worker calculate vector-graph CPU jobs.
|
|
# The Brain remains graph owner and validates every returned endpoint/value.
|
|
BRAIN_VECTOR_GRAPH_AGENT_OFFLOAD=false
|
|
BRAIN_VECTOR_GRAPH_AGENT_REQUIRED=false
|
|
BRAIN_VECTOR_GRAPH_AGENT_WAIT=2m
|
|
# AI-THINK consumes mathematical semantic_neighbor candidates before doing a new vector search.
|
|
BRAIN_THINKING_VECTOR_GUIDED=true
|
|
# Optional hard semantic 3D layout. Keep false unless you want an immediate full layout replacement.
|
|
BRAIN_VECTOR_GRAPH_LAYOUT=false
|
|
# Re-evaluate semantic proximity from the existing embeddings periodically without model calls.
|
|
BRAIN_VECTOR_GRAPH_REEVALUATE_INTERVAL=30m
|
|
# Gradually relax dense visual clouds toward the semantic layout instead of moving all nodes at once.
|
|
BRAIN_VECTOR_GRAPH_RELAX_LAYOUT=true
|
|
BRAIN_VECTOR_GRAPH_LAYOUT_RELAX_INTERVAL=2h
|
|
BRAIN_VECTOR_GRAPH_LAYOUT_BLEND=0.08
|
|
BRAIN_VECTOR_GRAPH_LAYOUT_MAX_SHIFT=0.035
|
|
BRAIN_CLUSTER_CANDIDATES_PER_ANCHOR=96
|
|
BRAIN_CLUSTER_ARTICLE_CANDIDATES=192
|
|
BRAIN_CLUSTER_REVIEW_EVIDENCE=8
|
|
BRAIN_CLUSTER_REVIEW_CONTEXT_CHARS=8000
|
|
# Bündelt thematisch kompatible Relationen eines THINK-Zyklus zu einem Artikeljob.
|
|
BRAIN_CLUSTER_ARTICLE_BATCHING=true
|
|
BRAIN_SIMILARITY_THRESHOLD=0.68
|
|
BRAIN_RELATION_THRESHOLD=0.72
|
|
BRAIN_TOP_K=8
|
|
BRAIN_MAX_CONTEXT_CHARS=16000
|
|
|
|
# Distributed Source-Agent inbox (Brain mode). Agent documents are queued outside
|
|
# the graph, embedded in small batches, and only promoted to candidate status when
|
|
# they match existing knowledge. Adaptive article research checks this inbox before SearXNG.
|
|
BRAIN_SOURCE_INBOX_ENABLED=true
|
|
BRAIN_SOURCE_INBOX_INTERVAL=30s
|
|
BRAIN_SOURCE_INBOX_BATCH_SIZE=12
|
|
BRAIN_SOURCE_INBOX_MIN_SIMILARITY=0.55
|
|
BRAIN_SOURCE_INBOX_MIN_PRIORITY=0.55
|
|
BRAIN_SOURCE_INBOX_NOVELTY_FLOOR=0.35
|
|
BRAIN_SOURCE_INBOX_MIN_RESULTS=2
|
|
# Freshness-sensitive article queries only reuse inbox documents newer than this.
|
|
BRAIN_SOURCE_INBOX_FRESH_MAX_AGE=168h
|
|
|
|
# Security/advisory exception: curated security candidates are not only cached.
|
|
# They enter a small proactive queue, are extracted by the article synthesis
|
|
# model, optionally enriched via direct web fetch/SearXNG, and can become an
|
|
# external security node before an article cites them.
|
|
BRAIN_SOURCE_INBOX_SECURITY_PROACTIVE=true
|
|
BRAIN_SOURCE_INBOX_SECURITY_BATCH_SIZE=2
|
|
BRAIN_SOURCE_INBOX_SECURITY_MIN_PRIORITY=0.58
|
|
BRAIN_SOURCE_INBOX_SECURITY_MIN_CONFIDENCE=0.72
|
|
BRAIN_SOURCE_INBOX_SECURITY_FETCH_MIN_CHARS=1800
|
|
BRAIN_SOURCE_INBOX_SECURITY_RESEARCH_RESULTS=3
|
|
|
|
# Optional controlled web research through your own SearXNG instance.
|
|
# Use the root URL or a URL ending in /search. Inside Docker, localhost points
|
|
# to the Brain container; use the SearXNG service name or host.docker.internal.
|
|
BRAIN_RESEARCH_ENABLED=false
|
|
SEARXNG_URL=
|
|
|
|
# Knowledge synthesis after a verified relation has formed a useful source cluster.
|
|
# Relation thinking always remains separate and only creates graph edges.
|
|
BRAIN_ARTICLE_SYNTHESIS_ENABLED=true
|
|
# Language tag for generated KB drafts (for example de-DE or en-US).
|
|
BRAIN_ARTICLE_LANGUAGE=de-DE
|
|
# Generate-then-review pipeline. Use a dedicated author model if it is available
|
|
# on at least one healthy Ollama pool node, e.g. gemma3:12b.
|
|
# The review model should be different from the author model when possible.
|
|
BRAIN_ARTICLE_SYNTHESIS_MODEL=qwen3:8b
|
|
BRAIN_ARTICLE_REVIEW_MODEL=qwen3:8b
|
|
# Number of targeted research + rewrite loops after the reviewer finds unsupported claims.
|
|
BRAIN_ARTICLE_REVIEW_REPAIR_ROUNDS=1
|
|
BRAIN_ARTICLE_MIN_SOURCES=3
|
|
BRAIN_ARTICLE_MAX_SOURCES=8
|
|
BRAIN_ARTICLE_MIN_PRODUCTION_RATIO=0.70
|
|
BRAIN_ARTICLE_MAX_GENERATION_DEPTH=2
|
|
BRAIN_ARTICLE_MIN_CONFIDENCE=0.74
|
|
BRAIN_ARTICLE_MIN_TEXT_CHARS=180
|
|
BRAIN_ARTICLE_MIN_ANSWER_CHARS=420
|
|
# Model-free pre-review. It scores depth, redundancy, source coverage and information density
|
|
# before the expensive Qwen claim/coverage review. Local CPU is the privacy-safe default;
|
|
# enable Agent offload only when that worker is trusted to receive article/source text.
|
|
BRAIN_ARTICLE_CPU_QUALITY_ENABLED=true
|
|
BRAIN_ARTICLE_CPU_QUALITY_AGENT_OFFLOAD=false
|
|
BRAIN_ARTICLE_CPU_QUALITY_AGENT_REQUIRED=false
|
|
BRAIN_ARTICLE_CPU_QUALITY_AGENT_WAIT=20s
|
|
# Web-Routing: auto = precise: always, clustered: adaptive.
|
|
# always = breite Vorabrecherche; adaptive = intern/Gemma zuerst und Web nur bei Aktualität/Evidenzlücke; review_only = Web nur auf Reviewer-Anforderung.
|
|
BRAIN_ARTICLE_RESEARCH_STRATEGY=auto
|
|
# Budget für initiale adaptive Webrunde (Aktualität oder Autor-Anforderung).
|
|
BRAIN_ARTICLE_ADAPTIVE_INITIAL_QUERIES=2
|
|
BRAIN_ARTICLE_ADAPTIVE_INITIAL_FETCH=3
|
|
# Maximal number of precise queries per research round.
|
|
BRAIN_ARTICLE_MAX_RESEARCH_QUERIES=6
|
|
# Raw SearXNG candidates per query.
|
|
BRAIN_ARTICLE_RESEARCH_RESULTS=12
|
|
# Upper research-planning budget. The active article pipeline performs one broad
|
|
# initial collection cycle; later targeted searches are driven by the reviewer.
|
|
BRAIN_ARTICLE_RESEARCH_ROUNDS=3
|
|
# Highest-ranked pages whose full content is downloaded.
|
|
BRAIN_ARTICLE_RESEARCH_FETCH_RESULTS=6
|
|
# Legacy/autonomous evidence mode and cached-material reuse still use these
|
|
# thresholds. The active generate-then-review article path itself fills its
|
|
# fetch budget from the highest-ranked SearXNG results without a semantic
|
|
# acceptance gate before drafting.
|
|
BRAIN_ARTICLE_RESEARCH_EXPLORATION_RESULTS=3
|
|
BRAIN_ARTICLE_RESEARCH_PREFETCH_MIN_RELEVANCE=0.25
|
|
BRAIN_ARTICLE_RESEARCH_MIN_RELEVANCE=0.55
|
|
BRAIN_ARTICLE_RESEARCH_MIN_QUALITY=0.35
|
|
BRAIN_ARTICLE_RESEARCH_PAGE_MAX_BYTES=2097152
|
|
BRAIN_ARTICLE_RESEARCH_PAGE_MAX_CHARS=14000
|
|
BRAIN_ARTICLE_RESEARCH_FETCH_TIMEOUT=20s
|
|
# Keep false unless private/intranet research URLs are intentionally trusted.
|
|
BRAIN_ARTICLE_RESEARCH_ALLOW_PRIVATE=false
|
|
|
|
# Research orchestration. All SearXNG requests, web fetches and Ollama calls
|
|
# share this bounded queue. With one Ollama node, 2 is a conservative default:
|
|
# one model request may run while one web request progresses.
|
|
BRAIN_RESEARCH_OLLAMA_MAX_INFLIGHT=2
|
|
BRAIN_RESEARCH_OLLAMA_QUEUE_SIZE=64
|
|
# Semantically equivalent research intents reuse a running/recent result bundle.
|
|
BRAIN_RESEARCH_DEDUPE_THRESHOLD=0.92
|
|
BRAIN_RESEARCH_DEDUPE_TTL=45m
|
|
# site: filters are always stripped at runtime. Research remains domain-open.
|
|
|
|
# Autonomous, persistent background research. Tasks are stored in graph.db and
|
|
# handed to Ollama asynchronously with low priority. Disabled by default.
|
|
BRAIN_AUTONOMOUS_RESEARCH_ENABLED=false
|
|
BRAIN_AUTONOMOUS_RESEARCH_IDLE_ONLY=true
|
|
BRAIN_AUTONOMOUS_RESEARCH_INTERVAL=30m
|
|
BRAIN_AUTONOMOUS_RESEARCH_TASKS_PER_CYCLE=1
|
|
BRAIN_AUTONOMOUS_RESEARCH_MAX_TASKS_PER_DAY=12
|
|
BRAIN_AUTONOMOUS_RESEARCH_MAX_QUERIES_PER_TASK=6
|
|
BRAIN_AUTONOMOUS_RESEARCH_MAX_PAGES_PER_TASK=8
|
|
BRAIN_AUTONOMOUS_RESEARCH_MAX_ROUNDS=3
|
|
BRAIN_AUTONOMOUS_RESEARCH_MIN_PRIORITY=0.65
|
|
BRAIN_AUTONOMOUS_RESEARCH_COOLDOWN=168h
|
|
BRAIN_AUTONOMOUS_RESEARCH_LEASE=45m
|
|
BRAIN_AUTONOMOUS_RESEARCH_MAX_ATTEMPTS=3
|
|
BRAIN_AUTONOMOUS_RESEARCH_QUERY_TRIGGERS=true
|
|
BRAIN_AUTONOMOUS_RESEARCH_OPPORTUNITY_LIMIT=8
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Agent mode only. These values are ignored in BRAIN_MODE=brain.
|
|
# Create/rotate the token under /source-agents.html in the Brain UI.
|
|
# A local JSON bootstrap file may be used instead of URL/ID/token envs.
|
|
# -----------------------------------------------------------------------------
|
|
# BRAIN_MODE=agent
|
|
# Docker Compose publishes the Agent status UI on BRAIN_AGENT_PORT (default 8092).
|
|
# BRAIN_AGENT_PORT=8092
|
|
# Use the Brain service name on a shared Docker network, a LAN/DNS address, or
|
|
# host.docker.internal:<BRAIN_PORT> for a separate Agent container on the same host.
|
|
# Never use 127.0.0.1/localhost for a separate container.
|
|
# BRAIN_AGENT_BRAIN_URL=https://brain.example.org
|
|
# BRAIN_AGENT_ID=security-news-01
|
|
# BRAIN_AGENT_TOKEN=brain_agent_xxxxxxxxx
|
|
# BRAIN_AGENT_CONFIG_FILE=
|
|
# BRAIN_AGENT_CONFIG_REFRESH=5m
|
|
# BRAIN_AGENT_HTTP_TIMEOUT=30s
|
|
# BRAIN_AGENT_CONCURRENCY=3
|
|
# BRAIN_AGENT_BATCH_SIZE=50
|
|
# Allow polling RFC1918/private URLs only for intentionally trusted intranet sources.
|
|
# BRAIN_AGENT_ALLOW_PRIVATE=false
|
|
# CPU compute worker. No Ollama/chat/embedding call is made for vector_graph or article_quality jobs;
|
|
# vector jobs receive existing float32 embeddings, article-quality jobs receive the already selected article/evidence text.
|
|
# BRAIN_AGENT_COMPUTE_ENABLED=true
|
|
# BRAIN_AGENT_COMPUTE_POLL_INTERVAL=5s
|
|
# BRAIN_AGENT_COMPUTE_MAX_BYTES=134217728
|
|
|
|
# Optional Docker-controller role. Keep disabled unless this Agent is intentionally
|
|
# trusted with the host Docker socket. Docker.sock is effectively host-root.
|
|
# BRAIN_AGENT_DOCKER_CONTROLLER_ENABLED=false
|
|
# BRAIN_AGENT_DOCKER_SOCKET=/var/run/docker.sock
|
|
# BRAIN_AGENT_DOCKER_COMPOSE_BINARY=docker
|
|
# BRAIN_AGENT_CONTROLLER_POLL_INTERVAL=5s
|
|
# BRAIN_AGENT_CONTROLLER_MAX_DURATION=15m
|