AI 系统的
数据层。

# an HNSW index over your keys — declared once,
# kept current by the write path
IDX.CREATE idx:sem ON PREFIX doc: FIELD vec TYPE vector KIND ann DIM 768 DISTANCE cosine M 16 EF 200

HSET doc:4410 title "Ada on pipelining" vec "<768 f32, little-endian>"

# nearest ten. no separate vector database, no sync job.
IDX.QUERY idx:sem KNN "<query vector>" LIMIT 10
-> 1) "doc:4410"
   2) "doc:9982"

embedding 由你给出；kevy 负责存它、索引它、检索它。引擎里没有模型，这是故意的。

Agent 记忆与 RAG →

IDX.CREATE idx:ft ON PREFIX doc: FIELD title TYPE str KIND text

IDX.QUERY idx:ft MATCH "pipelining"
-> 1) 1) "doc:1"
      2) "0.2877"          # BM25 score

# hybrid: fuse the text ranking with the vector ranking
IDX.QUERY HYBRID idx:ft MATCH "pipelining" idx:sem KNN "<vector>" LIMIT 20 RRFK 60

BM25，带中日韩分词，建在向量索引的同一批 key 上。

检索是怎么工作的 →

HSET order:1001 customer 881 status open  total 4400
HSET order:1002 customer 881 status paid  total 8400

IDX.CREATE idx:cust   ON PREFIX order: FIELD customer TYPE i64 KIND range
IDX.CREATE idx:status ON PREFIX order: FIELD status   TYPE str KIND range

# the read that would have been a SQL query
IDX.QUERY COMPOSE AND idx:cust EQ 881 idx:status EQ open
-> 1) "0"
   2) 1) 1) "order:1001"

带过滤的读始终是一次查表。没有查询计划器，没有扫描。

不靠数据库扛住读 →

# the answer, kept current by the WRITE path
VIEW.CREATE v:open881 QUERY ( AND idx:cust EQ 881 idx:status EQ open ) ORDER BY idx:cust

VIEW.QUERY v:open881
-> 1) "0"
   2) 1) "order:1001"  2) "881"

# reads never recompute it; writes keep it fresh

大多数应用向 ORM 真正索要的，就是这个。

物化视图 →

# a table is a declaration — compiled to named indexes, once
TABLE.DECLARE user PREFIX u: PK id COLUMN id str COLUMN name str COLUMN age i64 COLUMN dept str INDEX age range VALUES dept name ORDERPATH by_dept_age ON dept THEN age DESC

HSET u:1 id 1 name ada age 34 dept eng

# the ORDER BY dept, age DESC walk — one composite index, no planner
IDX.QUERY user.by_dept_age WHERE dept EQ eng LIMIT 20 FIELDS name age

类型化列、二级索引、复合 ORDER BY 路径——连你的 PG/MySQL schema 文件也能编译（kevy-cli sql compile）。没有运行期 SQL，没有 join：那些留在 Postgres。

单表服务型读 →

# kevy.toml — a RAM budget for the whole store
[tiering]
budget = "70%"               # or "4gb", or "auto"

# past the budget, the coldest values spill to a disk log
# and page back on access. a cold key is an ordinary key:
GET archive:2019:q3          # pays one disk read, same reply
TTL archive:2019:q3          # metadata answers from RAM
SCAN 0 MATCH archive:*       # sees cold keys — one key table

RAM 决定键的上限，磁盘决定数据的上限；AOF 持久化契约不动。v1 下沉字符串和 hash——list、set、stream 留在热层。

分层存储怎么工作 →

# tail every write from another process — or an agent.
# [feed] enabled = true in kevy.toml
FEED.SHARDS                 -> (integer) 16
FEED.TAIL 0                 -> 1) (integer) 1     # generation
                               2) (integer) 1     # offset
FEED.READ 0 1 0 COUNT 2     -> the writes themselves, replayable

偏移量可以续上。不用轮询，也不会漏。

变更流 →

# a 16-core server
cargo install kevy && kevy --port 6379

# inside your binary — no socket, no process
let db = Db::open("data/")?;
db.set(b"k", b"v", None)?;

# a browser tab — 481 KB, persists to OPFS
const db = await open({ persist: { name: "app" } });

# a microcontroller — no OS, no allocator
let mut store = Store::new_in(&mut arena);

四个地方，同一个引擎、同一批命令。

把 kevy 嵌进去 →