RAG & Search18ms ReflexFreemium / API Key
17,200 active downloads
RAG Relevance Scorer
Non-autoregressive relevance scorer for retrieved context documents
jevproxy // rag-chunk-filter (reflex kernel)
$npx jevproxy run rag-filter
JevProxy Intercept Resolved(Mode: DIRECT_REFLEX)
18ms|Cost: $0.0001|0 reasoning tokens burned
STDOUT • Tool Call Output:Status: 200 OK
agent.dispatch(filter_rag_chunks)
payload: { "query": { "type": "string", "description": "User search query" }, "chunks": { "type": "a...
✔ Decision short-circuited in 18ms without roundtrip to frontier LLM.
Upstream token bill saved: $0.0180 on this turn.
RAG Relevance Scorer
Compatible with Cursor, Claude Code, Windsurf, OpenCode
TRADITIONAL LLM CALL:UNOPTIMIZED
• Median Latency: 1620 ms
• Cost per Turn: $0.0180
• Mode: Full KV Cache Reload & TTFT Prefill
JEVPROXY REFLEX KERNEL:77x FASTER
• Median Latency: 18 ms
• Cost per Turn: $0.0001
• Accuracy: 98.5% deterministic
1-Click CLI Execution
Run this tool accelerated through the JevProxy gateway without manual wiring:
npx jevproxy run rag-filterOpenAI / Anthropic Tool Schema
JSON SpecificationPaste this schema into your agent tools definition or Cursor extensions:
{
"type": "function",
"function": {
"name": "filter_rag_chunks",
"description": "Score retrieved documents and prune irrelevant context chunks",
"parameters": {
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "User search query"
},
"chunks": {
"type": "array",
"items": {
"type": "string"
},
"description": "Retrieved text candidates"
},
"top_k": {
"type": "number",
"description": "Number of winning chunks to preserve"
}
},
"required": [
"query",
"chunks"
]
}
}
}Technical Architecture & Usage
Cutting Context Bloat Before Generation
Injecting 10 vector chunks into an agent prompt often wastes 8,000 tokens on irrelevant text. RAG Relevance Scorer filters down to the top 2 in 18.0ms.
Accelerate RAG Relevance Scorer with JevProxy
Get 5,000,000 free decision tokens. Eliminate the 3-second tool freeze in Cursor and Claude Code in under 60 seconds.