cortex.example.toml

# Cortex configuration. Copy to ~/.config/cortex/cortex.toml and edit.
# API keys are NEVER stored here -- they come from environment variables.

vault_path = "~/Documents/Obsidian"

# Obsidian vault name, used to build clickable obsidian:// citation links.
# Defaults to the vault directory's own name, which is what Obsidian uses
# unless you renamed the vault in the app.
# vault_name = "MyVault"
# data_dir  = "~/.local/share/cortex"

# Hard switch: when true, nothing leaves this machine regardless of policy.
local_only = false

exclude_globs = [".obsidian/**", ".trash/**", ".git/**", "**/node_modules/**"]

[embedding]
model = "qwen3-embedding:0.6b"   # ~1.5 GB. Best quality-per-GB at 16 GB.
dimensions = 1024
ollama_url = "http://localhost:11434"

[retrieval]
top_k = 8
dense_k = 30
fts_k = 30
graph_enabled = true
graph_hops = 1
rerank_enabled = true          # needs: pip install 'cortex-brain[rerank]'
rerank_model = "bge-reranker-v2-m3"   # or "minilm" -- 4x faster, lower ceiling
rerank_candidates = 30
rerank_idle_ttl = 600.0        # release reranker weights after 10 min idle
temporal_enabled = true        # resolve "last week" and return full coverage
expansion_enabled = true       # strip filler words, run extra lexical variants
memory_max_results = 2         # cap memory notes per result set (see below)
fusion_weights = { dense = 1.0, fts = 0.8, graph = 0.5 }

# Also index PDFs, HTML clippings, EPUBs and .txt files.
# Needs: pip install 'cortex-brain[documents]'
ingest_documents = false

[memory]
# Cortex writes session notes as plain Markdown into your vault, wikilinked to
# the notes they drew on -- so memory shows up in graph view and you can edit or
# delete it in Obsidian like anything else.
enabled = true
folder = "Memory"
# Write a note after every `ask`. Off by default: capturing every passing
# question fills the vault with noise. Use `cortex ask --remember` instead.
auto = false

[chunking]
target_tokens = 512      # 2026 consensus; semantic chunking underperforms this
overlap_ratio = 0.10
min_tokens = 64          # floor guards against micro-fragments
max_tokens = 900

[thermal]
# Tuned for a fanless 16 GB Air. Raise on a machine with fans.
boost_workers = 4
nominal_workers = 2
throttled_workers = 1
critical_workers = 0
throttle_speed_limit = 95
critical_speed_limit = 60
min_battery_for_backfill = 30
sample_interval = 10.0

# Providers patch the shipped defaults -- name an existing one to override
# single fields, or give base_url + model to add a new endpoint.
#
# [[providers]]
# name = "openrouter"
# zdr_enabled = true      # ONLY after enabling zero-data-retention on your account
#
# [[providers]]
# name = "cerebras"
# base_url = "https://api.cerebras.ai/v1"
# model = "llama3.1-8b"
# policy = "no_train"
# api_key_env = "CEREBRAS_API_KEY"
# tpd = 1000000
# priority = 15