AI Related
- AI (deepseek / opencode / chatgpt) used to generate parts of the code and comments
- Built-in embedding semantic search and LLM summaries (see the AI Commands section)
Initialization Notes for AI Agents / Scripts
By default no model is configured and nothing is vectorized — running --search directly will report "AI not configured" or "no vector index yet". AI should first run sip --config to confirm initialization; if config is missing run sip --init, if the index is missing run sip --index, if the model was changed run sip --reindex. Output is always UTF-8.
Retrieval Strategy (recommendations for agents)
- Confirm hits with full-text search first:
--grepis exact keyword matching (title/body/summary), no AI dependency, no threshold issues. The default is a safe snippet mode (each article only outputs "ID + title + hit count + ±50-char snippet", max 20 articles × 10 snippets), so large sources' bodies won't flood the context; if too many hits add--limit N, for structured results use--json, for a full article use--show <id> --json. - Then expand with semantic search:
--searchfinds "articles similar in meaning but different in wording" by semantic similarity; across all sources it's a full vector scan, which gets noticeably slower with large data — prefer--grep, use--searchonly when you truly need semantic expansion, combined with--feed <id>to limit sources and--thresholdto tune the threshold. - Vary the keywords: break the topic into 3–6 different keywords/phrases/synonyms/English originals, search each, then merge and dedupe.
- Watch the threshold: default 0.7. 0–2 results → lower to 0.5–0.6; lots of noise → raise to 0.75–0.8; local bge-m3 often lands at 0.5–0.6, so try 0.5; hits from full-text-fetched articles typically score 0.1–0.2 lower than title vectors, so you can lower the threshold and retry.
- Read the full text: use
sip --show <id> --json(AI should always pass--json; bare runs enter the full-screen reading UI); when a full-text cache exists, the JSON includes afulltextfield — use it first.
AI skill
The .opencode/skills/sip-rss directory in the source contains a skill you can hand directly to AI. You can also download sip-skill.zip directly from Releases (shipped alongside the single-file builds for each platform).