reflex-search / README.md
EdwardIrby's picture
docs: set initial public release version to v0.3.3
521d650 verified
|
Raw History Blame Contribute Delete
3.73 kB
metadata
license: apache-2.0
base_model: google/embeddinggemma-2
library_name: safetensors
pipeline_tag: text-classification
tags:
  - search
  - reranking
  - calibration
  - system-1
  - cpu-inference
  - embeddinggemma-2
  - mcp

youdotcom/reflex-search (v0.3.3)

Initial public release (v0.3.3) of reflex-search — a compact, CPU-native System 1 search decision and source-ranking model built on the text encoder of google/embeddinggemma-2 (d = 768, 24 text layers with hybrid 512-token sliding and global full attention). Vision and audio towers are removed so the text-only bf16 checkpoint occupies 517 MB on disk and runs in ~196 MB RSS (int8) / ~315 MB RSS (bf16) on commodity CPUs via s1_eg2cpu (adapted from undrgrnd-labs/hunch/eg2cpu).

It powers the reflex-search and reflex-extract MCP tools (/mcp/reflex):

  1. Gate 1 (Sub-query preflight): Scores parallel candidate sub-queries (noul_query_matches_intent, choice_best_query_for_intent) to filter off-topic hallucinations and rank over-budget queries before spending Search API calls.
  2. Gate 2 (Unified SERP + Knowledge ranking & routing): Scores web highlights and licensed knowledge facts (attribution + as_of) on one unified relevance scale (choice_best_search_hit + score_return_relevance), estimates counterfactual full-page read gain (noul_hit_worth_reading), and scores set-level coverage (score_evidence_coverage + choice_next_step in {answer, read, revise, stop}).
  3. Parallel Extract Intent Scoring: Scores BM25-windowed webpage and PDF passages against caller-supplied document_intent (score_extract_relevance on doc_kind = "extracted_doc").

Architecture & Artifact Layout

File Format Size Contents
model.safetensors BF16 517 MB Frozen text-only google/embeddinggemma-2 weights (language_model.*; zero vision/audio weights)
heads.safetensors F32 58 MB Rank-16 LoRA (layers [17, 20, 21, 22, 23] + embedding_projection, bypassed during prompt embedding for 0.0 retrieval drift) + ChoiceHead, NoulHead, and ScoreHead
heads.json JSON 9.8 KB Head dimensions, query templates, LoRA metadata, and post-hoc (T, b) calibration tables
calibration.json JSON 2.6 KB Single-shot untouched bar_locked holdout seal
tokenizer.json JSON 31 MB Gemma fast BPE tokenizer
config.json JSON 5.0 KB Backbone text configuration

Untouched Holdout Calibration (bar_locked, 6,155 Rows)

Trained and calibrated on leak-free search and counterfactual multi-rollout trajectories (0 root or query overlap with bar_locked and downstream evaluation benchmarks):

Cell Holdout Rows Proper Loss (vs. Base Rate) Accuracy / Conditional RPS Quantile ECE (Gate <= 0.045) Perturbation JS (Gate <= 0.005)
Choice:text 2,024 0.1250 < 0.1599 49.2% (5-way) 0.0307 0.00079
Noul:text 2,040 0.1564 < 0.2500 77.5% 0.0315 0.00130
Score:text 2,091 0.1185 < 0.3340 RPS 0.1554 < 0.2004 0.0166 0.00130
Combined 6,155 All 3/3 Heads Admitted — 0.0262 <= 0.040 0.00113 <= 0.005

License & Attribution

  • Base weights (model.safetensors, tokenizer.json, config.json) derive from google/embeddinggemma-2 under the Apache-2.0 license.
  • LoRA adapters, calibrated RLCD projection heads, and configuration metadata are released under Apache-2.0.