Laya for search and RAG
In retrieval pipelines Laya works best as a judge: is this passage relevant, does this evidence support the claim, should traversal continue. Projects use it to filter results and gate hallucinations. One published experiment found it is not a good general-purpose re-ranker.
88 projects
laya-demo
uulonger
Demonstrates typed Laya decisions for triage, moderation, guardrails, RAG filtering, and routing
codepilot
GorbunovS
Indexes code and provides MCP search tools with Laya-powered reranking
divya
RavaniRoshan
An India-focused market intelligence terminal with a fine-tuned Laya decision engine
laya-tools
teamswyg
Implements a Go-based local Laya router for code search and experimental model selection
chatbot-ai-company-python
special-snowflake
Ports a knowledge-base chatbot to Python and replaces its Jev model with Laya decision routing
LayaSpot
simo07d7PG
A search engine built using the Laya model
layagrep
asherepenko
Searches repositories by natural-language questions using a local Laya decision model
jgrep
rohit-burman
Provides a Go semantic grep CLI that scores text with a local Laya model
digital-twin
ShlokShah01
Builds a knowledge-grounded digital twin with GraphRAG, Laya, and optional voice
QQVibe
xzyj50609
Analyzes QQ chats locally for search, sentiment, intent, and participant profiles
laya-pipeline
trifleen
Uses Laya to route, compress, and check requests sent to local language models
jev-lens
xi029
A local workbench uses Laya, Jev, or Ollama to inspect RAG evidence decisions and replay policies
laya-seo
tomwugdgz
Audits websites for SEO using a local Laya model as a gate before escalating uncertain judgments to Jev
brain-openkit
hyeondata
Designs an Obsidian knowledge toolkit with a planned local Laya model for search and organization
recon-tactical-laya
orionhack99
Ranks web reconnaissance targets by risk using a local Laya model
github-stars
mokouliszt
Browses starred GitHub repositories and reranks search results with the Laya multilingual model
rag-benchmark
erendikmenn
A local Turkish RAG benchmark measures the contribution of Laya reranking
system1-rerank-bench
dchristopoulos
A benchmark compares Laya and other decision models with rerankers on BEIR retrieval datasets
jev-websearch-eval
anweat
An offline web-search evidence evaluation compares hosted Jev with local Laya and a rule baseline
laya-vs-reranker-benchmark
gowthxm07
Benchmarks Laya relevance filtering against cross-encoder reranking in RAG workflows
laya-model-research
RoyalMilkteaMaster
Research code evaluates Laya as a control model for multi-hop question-answering loops
mcp-router
Amantux
Routes MCP tools using a pluggable decision engine with Laya as its default local model
laya_explore
akshayaasd
Uses Laya to route queries among agents and tools in multi-agent and RAG workflows
jey_pro_v2
buibui2026
Builds and evaluates a Laya fine-tuning baseline for legal RAG judging
Filter and verify search results
zero-api-key-web-search searches the web without API keys and can add a local Laya relevance sieve:
pip install "zero-api-key-web-search[laya]" # pulls laya + torch
zero-search "react state management" --laya --laya-threshold 0.6 --json
zero-verify "Python 3.13 is the latest stable release" --laya --json
Without Laya installed, it falls back to a lexical verifier.
RAG claim checking
layaAsRagJudge runs the whole pipeline in the browser: chunking, MiniLM embeddings, top-k retrieval, then a Laya question asking whether the evidence supports the claim. An evaluation page scores 72 labelled claims.
GraphRAG
laya-jev-GraphRAG uses score, noul and choice questions for edge verification, entity disambiguation, A* traversal and a hallucination gate. It works across Neo4j, Memgraph, Apache AGE and Kùzu, with local Laya or cloud Jev.
Code retrieval
laya-codex ranks code for Claude Code with tree-sitter, BM25 and a Laya re-ranker (tindang/laya-code).
Long documents
laya-multilingual reads up to 8,192 tokens with max_len=8192. The core README reports 16 to 18 of 20 requests correct with up to about 4,000 tokens of text, and more variable results beyond that.
result = router.predict(long_document, questions, model="multilingual", max_len=8192)
Caveat: re-ranking
ai-experiments (nadeem4) re-ranked BM25 results for all 323 BEIR NFCorpus test queries. Jev improved nDCG@10 by +0.0347 and a MiniLM cross-encoder by +0.0204, while both Laya checkpoints ranked worse than doing nothing. Use Laya for yes/no relevance gates and claim checks. Measure before using it to reorder a ranked list.
More ways to use Laya