免费在 GitHub 公开

abi-superpower-sea

SEA (Sparse Evidence Attention) learning superpower. Evidence-augmented completion with task-aware scoring and adaptive modulation.

中文简介正在准备中。当前显示原文说明。

donaldfilimon/abi星标 19(GitHub 19 + cccskills 0)

实测评分未实测

查看安装方法

包含的文件（1）

SKILL.md2.1 KB

SKILL.md（原文）

安装前可以查看交给 AI 代理的指令内容。

ABI Superpower: SEA

Documents Sparse Evidence Attention learning as a superpower. There is no /abi-superpower-sea binary or slash command; use the real CLI path below.

Use

Run local evidence-augmented completion against a disposable durable store:

abi_sea_scratch="$(mktemp -d)"
ABI_WDBX_PATH="$abi_sea_scratch/wdbx" \
  ./target/debug/abi complete --learn "analyze document"
rm -rf "$abi_sea_scratch"

The task class and evidence limit are internal to this CLI path. There are no public --task or --evidence-limit flags, and the CLI has no separate learn, adaptive, or evidence subcommands. The MCP ai_learn tool accepts an optional evidence_limit capped at 100.

SEA Architecture

8 Signal Types: semantic, keyword, metadata, recency, authority, graph, contradiction, and task fit
Task-Aware Weighting: code repair, project recall, and benchmark review adjust the default weights; the other inferred task classes use the defaults
AdaptiveModulator: EMA weights (alpha=0.3) stored in WDBX key modulator:weights
Evidence Budget: deterministic selection bounded by record, token, cluster, and prompt-byte limits

Implementation

Maps to:

crates/abi-sea/src/learn_loop.rs - Core SEA algorithm
crates/abi-sea/src/evidence.rs - Evidence scoring and recall
crates/abi-sea/src/scorer.rs - Task-aware signal weights and ranking
crates/abi-sea/src/query_plan.rs - Query classification and retrieval plan
crates/abi-ai/src/constitution.rs - 6-principle constitutional audit
../wdbx/crates/abi-wdbx/src/store.rs - Persistent modulator weights

Runtime Boundary

SEA is linked into the Rust workspace; abi complete --learn selects the evidence path. Without --learn, completion uses the base route. The :memory: sentinel currently selects the disclosed no-store fallback, so SEA evidence smokes must set ABI_WDBX_PATH to a scratch durable directory and must never open the user's live ~/.abi store.

评价

还没有评价。欢迎分享你的使用感受。

无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

同一仓库的技能

abbey-system-prompt

免费

Maintain the Abbey / Aviva / Abi Discord system prompts in the Rust abbey-bot (`src/ask.rs`) and keep the hosted Grok bots (Grok Bot Abbey, Aviva Grok) in step with them. Use when the user says "improve grok bots", "abbey prompt", "persona prompt", "ask.rs", "system prompt", "tip floor", asks to change how Abbey/Aviva/Abi speak on Discord, or after an abbey-bot PR merges and the hosted bot's tip floor needs bumping. Runs as /abbey-system-prompt.

中文简介正在准备中。当前显示原文说明。