config.example.toml

# Copy this file to config.toml and adjust as needed.
# config.toml is gitignored (may contain machine-specific paths).

[storage]
# Path to the SQLite database. Relative paths resolve against the repo root.
db_path = "data/context.db"
# Path the VS Code extension appends events to (JSONL); ingested by the Python side.
vscode_jsonl = "data/vscode_events.jsonl"

[embeddings]
# Local Ollama. Make sure `ollama serve` is running and the model is pulled:
#   ollama pull nomic-embed-text
ollama_url = "http://localhost:11434"
model = "nomic-embed-text"
# nomic-embed-text produces 768-dim vectors. Used to sanity-check stored blobs.
dim = 768

[retrieval]
# Entries scoring below this cosine similarity are dropped entirely from results.
# This is the knob that stops unrelated context (the "irrelevant PDF") leaking in.
min_similarity = 0.45
# Default number of entries returned by get_user_context.
default_k = 8
# How strongly recent entries are favored. 0 = ignore recency, higher = stronger.
recency_strength = 0.15
# Half-life (in days) used for the recency factor at query time.
recency_half_life_days = 30.0

[decay]
# Multiplier applied to every entry's weight each time decay.py runs (daily cron).
# 0.98/day ≈ halves an un-reinforced entry's weight in ~34 days.
rate = 0.98
# Explicit facts decay slower (closer to 1.0 = barely decays).
explicit_rate = 0.997
# Entries whose weight drops below this are considered stale and pruned.
prune_below = 0.05

[ingest]
# Near-duplicate threshold: if a new entry is at least this similar to an
# existing one, reinforce the existing entry instead of inserting a new row.
dedupe_similarity = 0.92
# Minimum characters for an entry to be worth storing (skips noise like "ls").
min_chars = 12