📊
Knowledge Base Overview
Allectra's knowledge organized for AI-powered retrieval
Product Taxonomy Nodes
599
13 top-level categories, 4 levels deep
Products (SKUs)
6,631
All with EN/DE descriptions and pricing
Product Attributes
174
Technical properties across all categories
Knowledge Domains
7
Company, Product, Domain (incl. Cleaning & Cleanliness), Application, Customer, Ops, Process
Products with Properties
4,765
35,058 property assignments total
Priority Applications
Synchrotrons, Fusion
+Semiconductor as growth area, +Quantum as emerging
🎯 Design Principles
KI-ready from day one Hybrid retrieval (lexical + semantic) Bilingual EN/DE Layered: Taxonomy → Ontology → Knowledge Graph → Vector Index Incremental: no big-bang
The architecture captures not just what Allectra sells, but what Allectra knows: domain expertise in vacuum technology, application know-how for Big Science and industry, customer intelligence, and operational wisdom — all structured for AI agents to retrieve and reason over.
🏢
1. Company Knowledge
Unternehmenswissen — Who is Allectra?
- ▶1.1 Company Profile & History
- Founding story (2002, two physicist entrepreneurs)Gründungsgeschichte
- Mission: Trusted Partner for Ultra-High Vacuum
- Legal entities: Allectra GmbH, Allectra Limited, Manufacturing Hub
- Locations: Berlin (HQ), East Sussex (UK), Villingen-Schwenningen (DE)
- Revenue: ~€9M (2025), ~28 employees across group
- Exports to 30+ countries, offices/distributors in 15 countries
- ▶1.2 Manufacturing Capabilities
- CNC Lathe: Spinner TC600 (Ti/Al flanges, CM housings)
- CNC 3-Axis Milling: Spinner VC750-XL (PEEK plugs, flange milling)
- CNC 5-Axis Milling: Spinner U630 (chambers, Sub-D housings)
- Laser Welding: Alphalaser ALW450F 4-axis (4.5KW pulse / 450W CW)
- Automation: Robobox + Fanuc robot (Sub-D high-volume)
- >60,000 components supplied, >4,000 custom
- 24h turnaround for urgent beamline shutdownsdifferentiator
- ▶1.3 Quality & Compliance
- ISO 9001:2015 certified
- RoHS, REACH, Conflict Minerals compliant
- ASME-UD certified
- Leak testing: <5E-10 mbar L/s standard
- ▶1.4 Competitive Positioning
- Key differentiators: KAP301, CFS, Biallec™, SMPUSP
- Market: ~0.3-0.4% global, ~1.5-3% European
- Competitor landscape: VACOM, KJLC, MDC/ISI, Hositrad, MPF
🔧
2. Product Knowledge
Produktwissen — 6,631 products in 599 taxonomy nodes
- ▶Electrical Feedthroughs & ConnectorsElektrische Durchführungen14,874 props
- ▶Multipin (Sub-D, CM/M16, M12, USB, MS)
- Sub-D Feedthroughs (9, 15, 25, 37, 50 pin)
- Circular-Modular CM/M16
- M12 Feedthroughs
- USB Feedthroughsmodern
- Multipin MS Feedthroughs
- ▶Coaxial (SMA, SMB, SMP, BNC, MHV, SHV, ...)
- SMA (up to 27 GHz)
- SMP — smallest UHV coaxial, 4 per CF16USP
- BNC & BNC50, MHV & MHV50, SHV
- Triaxial, Type N, 7/16, Microdot
- ▶Power & High Voltage
- Power Feedthroughs (up to 20A continuous)
- High Voltage (up to 20kV DC)featured
- RF Power, Ceramic Breaks
- Thermocouple (Types C, E, J, K, N, R, S, T)
- ▶Cables & WiresKabel und Leitungen2,942 props
- KAP301 PFAS-Free Radiation Resistant KaptonFLAGSHIP
- Dipped Kapton Wire (Copper, Manganine)
- Ready-Made Vacuum Cables (Multipin, Coaxial)
- Thermocouple Wire (all TC types)
- PTFE / PEEK InsulatedPFAS risk
- Insulation Tubes, Kapton Foils & Tapes
- Air Side Cables, Tools & Accessories
- ▶Hardware (Flanges & Fittings)5,779 props
- CF Flanges (304L, 316L, 316LN, Titanium, BITAL)
- KF Flanges & Fittings
- ISO-K / ISO-F Flanges
- CFS Compact Flange System (proprietary)USP
- Glass-to-Metal Components
- Gaskets, Fasteners, Clamps
- Viewports1,996 props
- Motion in Vacuum1,234 props
- Valves966 props
- Fiber Optics1,341 props
- Gas/Liquid/Fluid Feedthroughs
- In-Vacuum Equipment (Sensors, Heaters, Fasteners)
- Safety (Pressure Burst Discs)
- Non-Magnetic / Titanium Components
- Miscellaneous (Glue, Anti-Seize, Sealant)
- ▶⭐ Proprietary Technologies
- KAP301: PFAS-free, 10&sup9; rad, 4K–300°C, 20kV DCfirst-mover
- CFS: 40% smaller/lighter than CF, compatiblepatented
- Biallec™: Ti-Al bimetal flanges, zero magnetismpatented
- SMP: Smallest UHV coaxial feedthrough (4 on CF16)unique
🧪
3. Domain Knowledge
Fachwissen Vakuumtechnik — The expertise that makes Allectra valuable
- ▶3.1 Vacuum FundamentalsRAG priority
- Pressure ranges: LV / MV / HV / UHV / XHV
- Gas kinetics: mean free path, molecular flow
- Outgassing theory & bake-out
- Leak detection (He leak testing)
- Units & conversions (mbar, Torr, Pa)
- ▶3.2 Flange & Sealing TechnologyRAG priority
- CF system: knife-edge seal, copper gasket, ISO 3669
- KF system: O-ring centering ring, ISO 2861
- ISO-K/ISO-F, JIS systems
- CFS (Allectra proprietary) design principles
- Gasket materials & bake-out compatibility
- ▶3.3 Feedthrough TechnologyRAG priority
- Ceramic-to-metal sealing (competitor technology: MDC/ISI, CeramTec, MPF)
- Allectra Glass Bead Sealing — proprietary UHV-tight conductor-to-housing sealPROPRIETARY
- ⚠ Process details INTERNAL ONLY — public: glass bead technique exists & is UHV-provenCONFIDENTIAL
- Signal type → feedthrough selection guide
- Specifications explained (voltage, current, frequency, pins)
- Custom design: multi-type combinations on single flange
- ▶3.4 Cable & Wiring in VacuumRAG priority
- Insulation materials: Kapton vs. PTFE vs. PEEK vs. bare
- PFAS regulation & KAP301 as alternativeregulatory
- Cable construction: stripping, soldering, crimping
- Selection guide: by signal, environment, application
- 3.5 Viewport Technology (materials, coatings, selection)
- 3.6 Motion & Manipulation in Vacuum
- 3.7 Materials Science for Vacuum (SS, Ti, Al, Cu, Ceramics, PEEK)
- ▶3.8 Cleaning & Cleanliness (Reinigung & Sauberkeit)NEWPRIORITY
- ▶3.8.1 Fundamentals & Terminology
- Cleaning vs. Cleanliness (removal vs. verified state)
- Contamination types: particulate, molecular, organic, inorganic
- Cross-contamination & recontamination prevention
- Design rules for cleanability (avoid crevices, blind holes, trapped volumes)
- ▶3.8.2 Procedures by MaterialAI-READY
- Stainless Steel: 5-step Allectra standard (degrease → neutralize → flush → DI rinse → dry)
- Al, Cu, Brass, Soldered: Ethanol ~100%, ultrasonic, 15 min (ignition hazard!)
- Ceramics, Glass, PEEK, PTFE: alkaline degreasing path
- Copper alloys: chromic/citric acid etch + hot DI water rinse
- ITER VQC1 chemical clean: 12-step process (solvent → vapor → wash → bake 300°C/24h)
- ▶3.8.3 Special Procedures
- Coated Fused Silica Viewports: darkroom, HEPA, spectroscopy-grade solvents, NO ultrasonic
- Electropolishing (VQC1): smooth surface for field emission — post-bake 450°C/24h
- Air Baking, Snow Cleaning (CO₂ blast), Vacuum Bellows
- ▶3.8.4 Cleanliness Grades & VerificationAI-READY
- ITER VQC classes: VQC1 (UHV, strictest) → VQC4 (moderate)
- ASML Molecular Grades: Grade 1 (EUV optics, XPS) → Grade 5 (naked eye)
- Verification: naked eye, UV-A (ISO 3059), RGA (outgassing), XPS (elemental)
- Key metrics: stain area (cm²/m²), outgassing rate (TPD160), HIO elements (at%)
- ▶3.8.5 Handling, Packing & Contamination Control
- Powder-free latex/nitrile gloves (NO colored gloves)
- No bare skin contact with vacuum surfaces after cleaning
- Continuous process: no breaks between chemical & rinse stages
- Packaging: Al foil + polyethylene bag, sealed under dry N₂
- ▶3.8.6 Regulatory & Customer RequirementsREGULATORY
- ITER Vacuum Handbook Appendix 13 (mandatory for ITER suppliers)
- ASML GSA 07 9510 (Cleanliness — Molecular, semiconductor reference)
- Allectra internal cleaning standard (5-step SOP + coated viewport SOP)
🔬
4. Application Knowledge
Anwendungswissen — What customers do with Allectra products
- ▶4.1 Big Science / Particle AcceleratorsPRIORITY
- Synchrotron beamlines: architecture, vacuum reqs, Allectra components
- References: Diamond Light Source, DESY, SOLEIL, PETRA IV
- Particle physics: CERN LHC/FCC, radiation-hard cabling
- FCC pipeline: €50-150M in passive components
- Neutron sources: ESS, ISIS
- ▶4.2 Fusion ResearchPRIORITY
- ITER: extreme conditions, ITR-19-004 Vacuum Handbook
- Wendelstein 7-X stellarator
- Emerging private fusion (CFS, TAE)
- ▶4.3 Semiconductor Manufacturinggrowth
- CVD/PVD/ALD, Etching, Ion implantation, EUV lithography
- PFAS impact on semiconductor fabs
- EU Chips Act: Infineon, GlobalFoundries, TSMC Dresden
- 4.4 Quantum Computing (trapped-ion UHV, neutral atom)emerging
- 4.5 Analytical Instruments (SEM, TEM, mass spec, surface analysis)
- 4.6 Other Cryogenics, Photonics, Space, Medical, Alternative Energy
- 4.7 Application → Product Mapping (decision support)agent-ready
👥
5. Customer & Market Knowledge
Kunden- und Marktwissen
- ▶5.1 Customer Segments
- Research Institutes & Universities
- Big Science Facilities (CERN, DESY, Diamond, ITER, ESS, FAIR)
- Industrial OEMs
- Distributors
- ▶5.2 Typical Customer Challengesagent FAQ
- "Which feedthrough for my application?"
- "Urgent component for beamline shutdown"
- "How to specify a custom multi-feedthrough flange?"
- "PFAS-free alternative to my PTFE cable?"
- "Can this viewport handle my laser power?"
- 5.3 Market Intelligence ($2.5B market, PFAS timeline, Big Science pipeline)
- 5.4 Competitor Intelligence
🛠
6. Operational Knowledge
Betriebswissen — Tips, Tricks, Selection Guides
- 6.1 Tips & Tricks (internal): assembly, stripping, bake-out, troubleshooting, cleaning SOPs, contamination prevention
- 6.2 Tips & Tricks (customer-facing): specification, installation, care, component cleaning, viewport cleaning (coated/uncoated), handling after cleaningagent content
- 6.3 Selection Guides: Feedthrough, Cable, Flange, Viewport, Gasketagent-ready
- 6.4 Standards Reference: ISO 3669, ISO 2861, IEC, MIL-STD, PFAS
⚙
7. Process Knowledge
Prozesswissen — How Allectra operates
- 7.1 Sales Process: quoting, lead time signals, custom engineering workflow
- 7.2 Engineering Process: custom design, material selection, testing
- 7.3 Quality Process: CAPA, inspection, leak testing, certificates
🏗
Layered Architecture
Five layers from source data to AI agents
Layer 5: AI Agent Interface
FutureConversational AI, advisory agents, chatbots
Layer 4: Vector Index + Hybrid Retrieval
Phase 3Embeddings + BM25, Reciprocal Rank Fusion
Layer 3: Semantic Layer
Phase 2-3Business definitions, reduces hallucination >50%
Layer 2: Knowledge Graph
Phase 2Product relationships, application mapping (Neo4j)
Layer 1: Ontology + Taxonomy
Phase 1599 nodes (SKOS) + domain model (OWL-Lite)
Layer 0: Source Data
ExistingERP, catalog, datasheets, expert knowledge
💡 Key Architectural Decisions
| Decision | Choice | Rationale |
|---|---|---|
| Graph Database | Neo4j | Mature ecosystem, GraphRAG support, community |
| Vector Database | Weaviate or pgvector | Hybrid search built-in |
| Taxonomy Standard | SKOS | W3C standard, lightweight, designed for this |
| Ontology Standard | OWL-Lite | Inference capability without over-complexity |
| Content Format | Markdown in Git | Version-controlled, developer-friendly, diffable |
| Retrieval Strategy | Hybrid (BM25 + Vector) | Part numbers need exact match; concepts need semantic |
| Chunking | Semantic + Proposition-based | Technical content needs coherent, self-contained chunks |
🧠
Inference Rules (Ableitungsregeln)
Formalized expert logic that derives new knowledge from existing facts
✨ Why Inference Rules Matter
Instead of storing every possible combination (6,600 products × 25 applications × 15 cleaning procedures = millions of entries), inference rules encode expert logic that derives the right answer dynamically. This transforms the Knowledge Base from a passive data store into a system that thinks like a vacuum engineer.
Where Rules Live in the Architecture
Layer 5: Agent Prompt Rules
Soft"If PFAS mentioned → proactively recommend KAP301"
Layer 3: Business Rules (YAML)
Core"Material + VacuumLevel + Temp → Flange + Gasket + Cleaning"
Layer 2: Graph Pattern Rules (Cypher)
Graph"Products on same flange → compatible_with (derived)"
Layer 1: OWL-Lite Axioms
Formal"UHV_Product ⊑ bakeable_min_200C" (formal logic)
Rule Types
| Type | Location | Formalism | Example |
|---|---|---|---|
| Ontology Axioms | Layer 1 | OWL-Lite | "Every CF product is bakeable to ≥200°C" |
| Graph Pattern Rules | Layer 2 | Cypher/SPARQL | "Products on same flange size are compatible" |
| Business Rules | Layer 3 | YAML catalog | "SS + VQC1 → ITER 12-step clean" |
| Agent Prompt Rules | Layer 5 | NL + conditions | "If fusion context → mention ITER cleaning" |
| Confidentiality Rules | All layers | Access flags | "Glass bead details = INTERNAL ONLY" |
Example Rules (YAML Format)
- ▶CLN-001: Material → Cleaning ProcedureBUSINESS RULE
- IF product.material IN [AISI 304L, 316L, 316LN] AND cleanliness ≠ VQC1
- THEN assign → ALLECTRA-5STEP-SS (degrease → neutralize → flush → DI rinse → dry)
- Confidence: 1.0 | Audience: internal, external, agent | Source: Cleaning-Procedure-Allectra.pdf
- ▶CLN-010: Application → Cleanliness Grade (Chain Rule)CHAIN RULE
- IF facility IN [ITER, ITER-like Tokamak] AND vacuum_zone = primary_vessel
- THEN required_cleanliness = VQC1
- TRIGGERS → CLN-005 (ITER 12-step) + CLN-011 (bake 300°C/24h) + HDL-001 (strict handling)
- Advisory: "All deviations require written ITER acceptance per Appendix 13"
- ▶SEL-001: Vacuum Level + Temp → Flange SelectionBUSINESS RULE
- IF vacuum_level IN [UHV, XHV] AND max_temperature ≥ 150°C
- THEN recommend CF flange | EXCLUDE KF, ISO-K (O-ring limit)
- IF non_magnetic required → use Titanium CF (Biallec™) instead of SS
- Cross-sell: copper gaskets + silver-plated bolts (anti-seize for bake-out)
- ▶ADV-001: PFAS Proactive AdvisoryAGENT RULE
- IF conversation mentions [PTFE cable, FEP cable, Teflon, fluoropolymer]
- THEN proactive advisory: "KAP301 is PFAS-free alternative, ECHA timeline..."
- Priority: HIGH | Links to: KAP301 product family, PFAS regulation timeline
- ▶CONF-001: Glass Bead Sealing — ConfidentialityCONFIDENTIAL
- IF query matches [glass bead seal*, Glasperlen*, sealing process detail*]
- PUBLIC: "Allectra uses proprietary glass bead sealing, UHV-proven"
- BLOCKED: process_parameters, temperature_profile, glass_composition
Execution Flow Example
"I need a feedthrough for our ITER tokamak beamline, what cleaning is required?"
1. Graph Query → Application(ITER) → requires → VQC1
2. Rule CLN-010 fires → assigns VQC1 → chains CLN-005, CLN-011, HDL-001
3. Rule CLN-005 fires → returns ITER 12-step procedure for SS
4. Rule HDL-001 fires → returns handling requirements (gloves, Al foil)
5. Agent ADV-005 fires → adds note about written ITER acceptance
6. Confidentiality check → no restricted content → full response
7. Response combines all conclusions into structured advisory
2. Rule CLN-010 fires → assigns VQC1 → chains CLN-005, CLN-011, HDL-001
3. Rule CLN-005 fires → returns ITER 12-step procedure for SS
4. Rule HDL-001 fires → returns handling requirements (gloves, Al foil)
5. Agent ADV-005 fires → adds note about written ITER acceptance
6. Confidentiality check → no restricted content → full response
7. Response combines all conclusions into structured advisory
Rule Governance
| Aspect | Recommendation |
|---|---|
| Storage | Git repository, one YAML file per rule domain |
| Versioning | Git history = full audit trail; semantic versioning |
| Review | PR review by domain expert + knowledge engineer |
| Testing | Unit tests: given conditions → expected conclusion |
| Conflicts | Higher confidence wins; if equal, more specific rule wins |
| Monitoring | Log rule fires per query; track hit rates & false positives |
| Expiry | Rules with last_verified >12 months → flagged for review |
🚀
Implementation Roadmap
Four phases from foundation to AI agents
Phase 1: Foundation
Months 1-3
- Formalize taxonomy as SKOS
- Define metadata schema
- Import 6,600 products
- Initial domain articles
- Basic hybrid search
Phase 2: Relationships
Months 4-6
- Product compatibility graph
- Application knowledge layer
- Selection guides
- Tips & tricks library
- Customer intelligence
Phase 3: AI Retrieval
Months 7-9
- GraphRAG implementation
- Semantic layer
- Query augmentation
- Evaluation metrics
- Domain embeddings
Phase 4: AI Agents
Months 10-12
- Conversational product advisor
- Role-based access
- Customer / internal / partner
- Feedback loops
- Continuous improvement
⚠
Anti-Patterns & Pitfalls
What to avoid when building the knowledge base
| Anti-Pattern | Description | Allectra Risk | Mitigation |
|---|---|---|---|
| Big Ball of Mud | Unstructured knowledge dump | HIGH | This architecture document defines the structure |
| Golden Hammer | Forcing everything into one tool | MEDIUM | Layered architecture with right tool per layer |
| Premature Over-Engineering | Complex graph before content exists | MEDIUM | Phase-based: start with taxonomy + search |
| Knowledge Debt | "Truck factor of one" — only one person knows | HIGH | Systematic capture in knowledge base |
| Wikiphobia | Barriers too high for contributors | MEDIUM | Markdown + templates, low friction |
| Stale Knowledge | No maintenance process | HIGH | Freshness dates, review cycles, quality gates |
| Vector-Only Search | Ignoring exact match needs | MEDIUM | Hybrid retrieval (BM25 for part numbers + vectors for concepts) |
| Ignoring Bilingual | EN-only loses DE-speaking customers | MEDIUM | Bilingual from day one, DE labels on all entities |
✅ Success Factors
Start simple, evolve incrementally Domain ontology investment (vacuum expertise) Metadata enforcement from day one Hybrid retrieval (exact + semantic) Dedicated maintenance budget Evaluation-driven iteration Treat as product, not project