Structured data layer: a six-library knowledge graph
The middle layer is not a loose yellow page. Materials / formulas / standards / papers / patents / suppliers — six tables interlinked by foreign keys and semantic similarity — form a knowledge graph that always has a next step for engineers who walk in.
Four data principles
Material pages show related formulas / recommended suppliers / downloads; formula pages link back to material entries; paper pages show related papers; patent pages surface peer patents. Every page is an entry to the other five libraries.
Each material parameter will carry a source_type (manufacturer TDS / third-party test / AI inference). Papers link out to DOI, patents link to Google Patents.
Cron runs at 2am UTC daily, pulling from CrossRef + OpenAlex + Google Patents + RSS. Ingest → translate → Chinese commentary → sitemap refresh → multi-engine push.
Beyond the automatic cron, companies can claim their own supplier listing and fill in details; engineers can submit corrections.
Current volume & structure
Covers resin / glass fiber (yarn / mat / fabric) / carbon / basalt / aramid / bio-based / core / gelcoat / auxiliary / composite product. Each carries mechanical, thermal, and physical properties in structured form.
Grouped by 7 processes (hand lay-up / filament winding / pultrusion / compression molding / RTM / vacuum infusion / 3D printing). Each entry includes resin system, reinforcement, auxiliary, process parameters, expected properties, operating tips, safety notes.
China GB / HG / JC, International ISO, American ASTM, European EN, German DIN. Organized by process and product. AI clause commentary on the roadmap.
Aggregated from CrossRef + OpenAlex. Covers mechanics, processing, fiber reinforcement, resin matrix, durability, recycling. Each with AI Chinese commentary.
Aggregated from Google Patents — CN / US / EP / JP / KR. Covers pultrusion / RTM / HP-RTM / thermoplastic / window/door / automotive / wind. Commentary backfill in progress.
Real composites industry firms. Claimed companies carry contact emails that receive RFQs directly. Includes major Chinese composites leaders (Jushi, Zhongfu Shenying, Taihe, Huachang, etc.).
Stack
ap-southeast-1 region, usage-based billing. Schema: 14 tables + 5 enums. Drizzle ORM for type-safe queries.
Each entry embedded via OpenAI text-embedding-3-small (1536-dim) and stored in knowledge_chunks with HNSW index for millisecond semantic search.
sources.ts provides unified interface to CrossRef / OpenAlex / Google Patents / RSS. Daily cron rotates through query-pool topics. Dedupes by DOI / publication number.
Gemini 2.5 Flash does translate and commentary; Claude Haiku via OpenRouter fallback on high failure. Pipeline is idempotent and re-runnable.
Entry points
Every library has its own list page with multi-dimensional filters (category / brand / country / year / language / process), search, and compare.
Material pages show properties table + recommended suppliers + related formulas + downloads. Paper / patent pages show Chinese commentary + related entries + external source link.
Every [#N] in an AI answer links to a data-layer entry. Data is not isolated — it is evidence for AI answers.
Next-year expansion
Expand data itself to DE / JA / RU / ES / PT. Doubles overseas reach.
2027 H1
Every material parameter gets source_type (TDS / 3rd-party / AI-inferred) + confidence (0-1). UI shows Verified / Claimed / Estimated badges.
Q4 2026
Upgrade key parameters (tensile / modulus / Tg / water-absorption) from single values to temperature × parameter series. Add S-N fatigue, creep, and water-absorption curves.
2027 H1
Open REST API to ERP / PLM / CAE vendors, per-call billing. Become the industry data substrate.
Q1 2027
Grow materials from 4,341 to 100,000-scale, on par with MatWeb. Add more grades, batches, and temperature-dependent properties.
All of 2027
Ongoing iteration, suggestions welcome.