NeuralChainX
Get in touch
Technology 01 Retrieval engine

Anneal Model

A domain-specialized embedding and retrieval engine for semiconductors.

Six source streams are aligned into one vector space, and questions and documents are measured on the same scale to find the evidence. Retrieval accuracy becomes answer reliability, so this engine sets the ceiling for the whole product.

01 — Pipeline

Data takes one path from arrival to decision artifact.

Six source streams enter the same pipeline. Anneal holds the retrieval scale at the middle of it, and reasoning runs on top of that.

00 SOURCE STREAMS
Patents Papers · Research Companies · Supply Chain Equipment · Process Logs Internal Documents Decision History
01 INGESTION Multi-source ingestion Notation and units are normalized to one standard.
02 STRUCTURED EVIDENCE Structured extraction Entity resolution and a citation graph turn source text into one comparable row.
03 ANNEAL ENGINE Embedding · vector index · retrieval Chunked documents and the question are measured on the same scale to pick the nearest evidence.
04 ORCHESTRATION Multi-agent reasoning Role-specialized agents check the evidence they were handed, through an LLM.
05 DECISION ARTIFACTS Decision artifacts Comparison tables and decision packs go out, with a source on every sentence.
↺ Judgments that clear review return to the corpus and feed the next training run
02 — Retrieval engine

Questions and documents are measured on the same scale.

Learning from domain data and the review history of its users, it sets the distance between semiconductor terms and concepts the way engineers in the field read them. The finer that scale, the better the evidence on the first screen.

ANNEAL · DOMAIN-SPECIALIZED RETRIEVAL ENGINE
question near documents far documents
INDEX MODELS
anneal-1-alpha granite-r2-base

A dedicated index model per corpus.

WHY IT MATTERS

Everyone calls the same LLM. What that LLM reads is decided by Anneal.

Patents, papers, company data and internal documents are searched together by similar technical content.

03 — What compounds

Each step to the right is harder to copy.

The raw material is public. What follows it — the processing, the experience, and the assets that accumulate — opens a gap that widens with time.

01 · MULTI-SOURCE CORPUS

Raw material

Six public and internal source streams collected through one pipeline. The collection is reproducible at any time.

02 · STRUCTURED EVIDENCE

Processing

Source text is restructured into one comparable row. Conditions, measured values and the location of the evidence sit in the same table.

03 · AGENTIC WORKSPACE

Experience

Work is delegated to role-specialized agents in a problem-scoped workspace. It fits the organization more closely the more it is used.

04 · COMPOUNDING CORPUS

Lock-in

Internal material and decision history join under tenant isolation. As artifacts accumulate they become an asset only that customer holds.