Repository navigation
Releases
Releases · TsukimioRinka/agg-search
v0.4.0 — agent x domain catalog, tuned profiles
agg-search-install --catalog+ catalog.json + docs/catalog.md: per-agent and per-domain entry points (Cursor/VS Code deeplinks, Claude Code/Codex commands).- Profiles wait for their primary engine in fast mode and blend Jina with RRF (academic 0.4, code 0.6): precision@3 academic 4/9→8/9, code 4/9→7/9.
- docs/agent-tests.md: isolated CLI checks for MCP Inspector, Codex, OpenCode, Claude Code, Gemini CLI, Antigravity CLI.
v0.3.0 — domain profiles, keyless domain engines, feedback loop
--profile academic|code|newspresets engines, RRF weights, recency and snippet length; MCPprofileandAGG_SEARCH_PROFILEmake it per-client.- Keyless engines
openalex,hn,stack. search+fetchfeedback feedsagg-search-tune, which writes clamped engine-weight multipliers (hash/counter only,--no-learnopts out).
v0.2.1 — injection hardening, installer check
- Model-facing text drops zero-width and bidi controls; MCP marks page text untrusted.
agg-search-install --checkvalidates config, entry and runtime.fetchtext_limit capped at 200k chars in CLI and MCP schema.
v0.2.0 — agent-facing MCP surface, HTTP transport, filters
MCP 2025-06-18 (version negotiation, structured output, annotations, hard timeouts, fast default); streamable HTTP transport (agg-search-http); per-client installer (agg-search-install, 13 clients); domain include/exclude filters; --show-config; SSRF guard and page cache for fetch; py.typed and a stdlib-only CI guard.
v0.1.2 — registry packaging and version reporting
- README carries the
mcp-namemarker the official MCP registry uses to verify package ownership. initializenow reports the installed version instead of a hardcoded0.1.0.server.jsontargets 0.1.2; the PyPI job inpublish.ymlis skipped untilPYPI_PUBLISHis set.
v0.1.0 — aggregate search, MCP server, tiered fetcher
First public release. Standard library only on the host; every engine is optional.
What ships
- CLI —
agg-search "query": parallel recall across exa, brave, serper (Google SERP), deepseek, firecrawl, anysearch, tavily and searxng. RRF fusion into the top 15, then Jina Reranker v3.5 (listwise, fallback v3 → v2) down to the final top N. - Resilience — 5-minute result cache, failure-class cooldowns (quota 1h / anti-bot 10min / timeout 90s), engine health in
--json, missing keys skipped up front. - Extras —
--expand N(RAG-Fusion via local ollama),--answer(cited summary; local or DeepSeek),--fast(first sources win),--json. - Tiered fetch —
crawl-fetch: DeepSeek Fetch API first (no container), then a static → dynamic → stealth container ladder. - MCP server —
agg-search-mcp(stdio JSON-RPC):searchandfetchtools,initialize/tools/list/tools/call/ping.
Install
git clone https://github.com/TsukimioRinka/agg-search
cd agg-search && python3 -m pip install -e .Keys come from the environment only; see docs/environment.md.
Verified
17 offline tests (no network, no quota) across Python 3.10 / 3.11 / 3.12.
Known limitation
Fetch tiers 1–3 need the external crawl-agent container image, which is not part of this repository. --tier deepseek works without it.