wd041216-bit/free-web-search-ultimate

⚡ Jev-Powered ⚡ — free neural search & evidence verification for AI agents.
Zero API keys. MCP-ready. Batch-filter, rerank and verify evidence locally at 100+ decisions/sec with an open Jev-class decision model.
---
Why Zero-API-Key Web Search (Jev-Powered)
An agent that searches the raw web pays for everything it reads: context tokens, latency, and attention spread over SEO noise. Zero-API-Key Web Search puts a **sieve between search and the agent**:
1. **Search** the web with zero API keys (DuckDuckGo by default, self-hosted SearXNG for cross-validation, optional Bright Data for production SERP + Web Unlocker).
2. **Sieve** the results through Laya — an open-weight (Apache-2.0), non-autoregressive *decision model* that scores every result for relevance in a **single forward pass (~35 ms on GPU, 100–330 questions/sec batched)** with calibrated probabilities. It never generates text, so there is nothing to parse and nothing to hallucinate.
3. **Verify** claims with probabilistic stance classification (support / conflict / neutral per source) instead of pure keyword matching, then emit citation-ready evidence reports.
Every stage degrades gracefully: no Laya installed? The pipeline falls back to the lexical heuristic verifier and still works — Laya is an accelerant, not a dependency.
30-Second Setup
Add the Sieve (Optional, ~800 MB, Fully Local)
On Apple Silicon it runs on MPS; on NVIDIA GPUs on CUDA; anywhere on CPU (slower but works).
Behind a network that can't reach Hugging Face? export HF_ENDPOINT=https://hf-mirror.com.
What Laya does inside the pipeline
| Stage | Question asked (one forward pass per item) | Effect |
| --- | --- | --- |
| --laya on search/context | *"Is this result relevant and informative for the query?"* | drops noise **before** it burns context tokens |
| --laya on verify/report | *"Does this text support the claim?"* + *"Does it state or imply the claim is false?"* | replaces regex conflict markers and keyword-overlap cutoffs with calibrated probabilities |
The verification blend becomes 0.55 × neural stance + 0.35 × source quality + 0.10 × freshness (vs. the lexical model's 0.45 × keyword overlap + …). The result payload keeps the full lexical sidecar so you can compare both classifiers on your own data.
Checkpoints
| Subfolder | Backbone | Best at |
| --- | --- | --- |
| root (default) | ModernBERT-large 421M | English text |
| multilingual | mmBERT-base 322M | 100+ languages, ~2× faster |
| typed-decisions | ModernBERT-large 421M | workflow-tuned decisions |
MCP Server (8 tools)
search_web, llm_context, browse_page, verify_claim, evidence_report, list_providers, clear_cache, setup_providers — the first four accept "laya": true and "laya_threshold" so a Claude/Cursor/any-MCP agent can opt into the sieve per call.
Provider Paths: Free to Production
**Path 1 — Free (zero configuration).** DuckDuckGo, no account:
**Path 2 — Free cross-validated.** Self-hosted SearXNG for dual-provider corroboration (see docs/searxng-self-hosted.md):
**Path 3 — Production SERP.** Bright Data for 7 engines (Google, Bing, DuckDuckGo, Yandex, Baidu, Yahoo, Naver), geo-targeting, structured results: