Zwei RAG-Architekturen über demselben Korpus aus 40 Bewerbungs-PDFs befragen: proRAG sucht über Vektoren, lightRAG über einen Graphen. Erwartete Dauer: ca. 45 Minuten.
Beide Systeme arbeiten auf demselben Korpus — dieselben Lebensläufe, zweimal indiziert. Die Datenbanken teilen sie sich NICHT: jedes System hat seinen eigenen Index und antwortet deshalb anders.
proRAG und lightRAG fragen beide nach diesen Zugangsdaten. Benutzername = Rechnername: Roberto, Giovanni, Giulia, Elena (manchmal MITSM):
# Login (proRAG + lightRAG) Username: •••••••• Password: ••••••••
Die 40 Lebensläufe, jede Zeile verlinkt ihr PDF.
Alle als ZIP herunterladen (cv.zip)
| Vorname | Nachname | Rolle | Klasse | Sprachen | Skills | |
|---|---|---|---|---|---|---|
| Grace | Adeyemi | HR Manager / Recruiter / Corporate Trainer | B, D, DE | Italian Native, English C2 | ·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·Emergency response: evacuation procedures, fire safety and incident coordination·Change management: rollout planning, training and adoption support·Team leadership: shift coordination, staff scheduling, onboarding and mentoring new hires·Cybersecurity awareness: access control, phishing prevention and data handling·Content creation: copywriting, social media and photography·Multilingual communication: fluent cross-language written and spoken support· | |
| Kwame | Adeyemi | Security Officer / Facilities Manager / Youth Football Coach | B, BE, D | English Native, Arabic B2, Polish A1, Spanish C2 | ·Data analysis: SQL, Python (pandas), Power BI and Tableau·Blueprint and technical drawing interpretation·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Budget ownership: cost control, forecasting and variance reporting·Clinical documentation: care plans, handovers and SOAP notes·Security operations: access control, CCTV monitoring, patrols and incident response· | |
| Fatima | Al-Rashid | Journalist / Podcast Producer / English Teacher | B, C | Arabic Native, Italian B2, German B2 | ·Security operations: access control, CCTV monitoring, patrols and incident response·Multilingual communication: fluent cross-language written and spoken support·Data entry and records management: accuracy, GDPR-compliant handling·Customer service: complaint handling, expectation setting and follow-up·Budget ownership: cost control, forecasting and variance reporting·Negotiation: supplier terms, contracts and conflict resolution·Performance management: goal setting, appraisals and constructive feedback· | |
| Henrik | Andersen | Pilot / Flight Instructor / Drone Operator | B, BE | English Native, German B1, Spanish C1, French C1 | ·Project management: planning, milestone tracking, stakeholder updates and delivery·Time management and prioritisation under deadline pressure·Quality assurance: inspections, audits and compliance checks·Adaptability: thriving in fast-changing and multi-tasking environments·Negotiation: supplier terms, contracts and conflict resolution·Web development: HTML, CSS, JavaScript and responsive design·Customer service: complaint handling, expectation setting and follow-up· | |
| Niklas | Berg | Carpenter / Furniture Maker / Site Foreman | A, B | German Native, Polish B1 | ·Presentation and public speaking to technical and non-technical audiences·Stock management systems: SAP EWM, barcode scanners and RFID·Spreadsheets: Excel/Power Query, pivot tables, formulas and dashboards·Electrical installation: wiring, testing and fault diagnosis·Performance management: goal setting, appraisals and constructive feedback·Procurement and invoicing: purchase orders, reconciliation and vendor records·Technical writing: documentation, SOPs and knowledge-base articles· | |
| Marcus | Bergmann | Sales Manager / Account Executive / Real Estate Agent | A1, B | English Native, French B2 | ·Attention to detail and accuracy in repetitive tasks·Stakeholder communication: de-escalation, clear written and verbal updates, reporting·Customer service: complaint handling, expectation setting and follow-up·Multilingual communication: fluent cross-language written and spoken support·Manual handling and ergonomics: safe lifting and workstation assessment·Digital tools: Microsoft 365, scheduling software, CRM and ticketing systems·Supply chain coordination: procurement, supplier management and delivery scheduling· | |
| Lukas | Brandt | Warehouse Supervisor / Forklift Instructor / Retail Manager | B | English Native, Italian A1, Spanish B2, Arabic B2 | ·Financial reporting: P&L, cash-flow analysis and month-end close·Facilities management: contractor coordination, planned and reactive maintenance·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Emergency response: evacuation procedures, fire safety and incident coordination·Multilingual communication: fluent cross-language written and spoken support·Supply chain coordination: procurement, supplier management and delivery scheduling·Customer service: complaint handling, expectation setting and follow-up· | |
| Isabella | Conti | Fashion Designer / Tailor / Retail Buyer | B, BE | Italian Native, Arabic A2 | ·UX/UI design: wireframing, prototyping and user research·Time management and prioritisation under deadline pressure·Financial reporting: P&L, cash-flow analysis and month-end close·Electrical installation: wiring, testing and fault diagnosis·Customer service: complaint handling, expectation setting and follow-up·Content creation: copywriting, social media and photography·Data entry and records management: accuracy, GDPR-compliant handling· | |
| Aisha | Diallo | HR Coordinator / Translator / Childcare Assistant | A1, B, C1, C1E, D1 | Arabic Native, English C2 | ·Attention to detail and accuracy in repetitive tasks·Procurement and invoicing: purchase orders, reconciliation and vendor records·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Time management and prioritisation under deadline pressure·Spreadsheets: Excel/Power Query, pivot tables, formulas and dashboards·Data entry and records management: accuracy, GDPR-compliant handling·Content creation: copywriting, social media and photography· | |
| Sophie | Dubois | Physiotherapist / Sports Masseur / Wellness Coach | B, C, D, DE | Italian Native, Polish B1, French C1, Spanish C2 | ·Financial reporting: P&L, cash-flow analysis and month-end close·Health & safety compliance: risk assessments, incident reporting, first aid trained·Electrical installation: wiring, testing and fault diagnosis·Presentation and public speaking to technical and non-technical audiences·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Plumbing and HVAC basics: repairs, servicing and inspections·Blueprint and technical drawing interpretation· | |
| Ahmed | Farouk | Data Engineer / Analytics Consultant / Coding Tutor | B | French Native, English C2, Polish B1 | ·Event management: budgeting, vendor sourcing, logistics and on-site delivery·Project management: planning, milestone tracking, stakeholder updates and delivery·Rehabilitation and mobility support: exercises, aids and progress tracking·Electrical installation: wiring, testing and fault diagnosis·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Adaptability: thriving in fast-changing and multi-tasking environments·Performance management: goal setting, appraisals and constructive feedback· | |
| Marco | Fernandez | IT Project Manager / Barista Trainer / Logistics Coordinator | B, BE, C1, C1E | German Native, English B1, Italian B1, Arabic C2 | ·Negotiation: supplier terms, contracts and conflict resolution·Clinical documentation: care plans, handovers and SOAP notes·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·Web development: HTML, CSS, JavaScript and responsive design·Medication administration and dosage management·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Blueprint and technical drawing interpretation· | |
| Lena | Fischer | Primary Teacher / Tutor / Museum Guide | A1, B | French Native, Spanish B2, Arabic A2, Italian B2 | ·Rehabilitation and mobility support: exercises, aids and progress tracking·Procurement and invoicing: purchase orders, reconciliation and vendor records·Cybersecurity awareness: access control, phishing prevention and data handling·Customer service: complaint handling, expectation setting and follow-up·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Facilities management: contractor coordination, planned and reactive maintenance·Supply chain coordination: procurement, supplier management and delivery scheduling· | |
| Omar | Haddad | Civil Engineer / Site Manager / CAD Technician | B | Spanish Native, Polish A2 | ·Adaptability: thriving in fast-changing and multi-tasking environments·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Electrical installation: wiring, testing and fault diagnosis·Cybersecurity awareness: access control, phishing prevention and data handling·Content creation: copywriting, social media and photography·Automation: scripting, workflow tools and no-code integrations·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation· | |
| Leila | Hassan | Social Worker / Youth Counsellor / Community Organiser | A2, B, C1 | Polish Native, French C2, Spanish B1, Italian A1 | ·Rehabilitation and mobility support: exercises, aids and progress tracking·Team leadership: shift coordination, staff scheduling, onboarding and mentoring new hires·Manual handling and ergonomics: safe lifting and workstation assessment·Administrative support: diary management, minute-taking and correspondence·Presentation and public speaking to technical and non-technical audiences·IT support: troubleshooting, hardware setup and user account administration·Blueprint and technical drawing interpretation· | |
| Nadia | Hoffmann | Registered Nurse / Palliative Care Specialist / Clinical Educator | B, BE, C1, C1E | French Native, German C1, Arabic A2 | ·Cybersecurity awareness: access control, phishing prevention and data handling·Adaptability: thriving in fast-changing and multi-tasking environments·Quality assurance: inspections, audits and compliance checks·Security operations: access control, CCTV monitoring, patrols and incident response·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·Facilities management: contractor coordination, planned and reactive maintenance· | |
| Pavel | Horák | Welder / Metal Fabricator / Workshop Supervisor | B | Polish Native, German A2, Spanish A2, Italian C2 | ·Technical writing: documentation, SOPs and knowledge-base articles·Content creation: copywriting, social media and photography·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Web development: HTML, CSS, JavaScript and responsive design·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·Adaptability: thriving in fast-changing and multi-tasking environments· | |
| Hannah | Kim | Graphic Designer / Photographer / Social Media Manager | B, C1 | Spanish Native, Italian C2 | ·Time management and prioritisation under deadline pressure·Negotiation: supplier terms, contracts and conflict resolution·IT support: troubleshooting, hardware setup and user account administration·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Data entry and records management: accuracy, GDPR-compliant handling·Team leadership: shift coordination, staff scheduling, onboarding and mentoring new hires·Manual handling and ergonomics: safe lifting and workstation assessment· | |
| Anna | Kowalczyk | Dental Hygienist / Receptionist / Pilates Instructor | B, BE, D, DE | French Native, English A1, Italian B1 | ·Performance management: goal setting, appraisals and constructive feedback·Negotiation: supplier terms, contracts and conflict resolution·Clinical documentation: care plans, handovers and SOAP notes·Graphic design: Adobe Creative Suite, branding and layout·Rehabilitation and mobility support: exercises, aids and progress tracking·Facilities management: contractor coordination, planned and reactive maintenance·Digital tools: Microsoft 365, scheduling software, CRM and ticketing systems· | |
| Tobias | Krause | Plumber / Heating Engineer / Handyman | AM, B | Italian Native, German C1 | ·Time management and prioritisation under deadline pressure·Technical writing: documentation, SOPs and knowledge-base articles·Clinical documentation: care plans, handovers and SOAP notes·Electrical installation: wiring, testing and fault diagnosis·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Budget ownership: cost control, forecasting and variance reporting·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation· | |
| Ingrid | Larsen | Architect / Interior Designer / Ceramics Teacher | A, B | German Native, Italian A2, Spanish B1, Polish C1 | ·Problem solving: root-cause analysis and practical decision making·Emergency response: evacuation procedures, fire safety and incident coordination·Team leadership: shift coordination, staff scheduling, onboarding and mentoring new hires·Time management and prioritisation under deadline pressure·Rehabilitation and mobility support: exercises, aids and progress tracking·Stock management systems: SAP EWM, barcode scanners and RFID·Clinical documentation: care plans, handovers and SOAP notes· | |
| Emma | Lindqvist | Veterinary Nurse / Animal Trainer / Pet Groomer | AM, B | German Native, English C2 | ·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Team leadership: shift coordination, staff scheduling, onboarding and mentoring new hires·Administrative support: diary management, minute-taking and correspondence·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Graphic design: Adobe Creative Suite, branding and layout·IT support: troubleshooting, hardware setup and user account administration·Customer service: complaint handling, expectation setting and follow-up· | |
| Mei | Ling Chen | Pharmacist / Research Assistant / Yoga Teacher | A1, B, BE, D | Polish Native, German C1, English C2, Spanish A1 | ·Content creation: copywriting, social media and photography·Customer service: complaint handling, expectation setting and follow-up·Graphic design: Adobe Creative Suite, branding and layout·Performance management: goal setting, appraisals and constructive feedback·Stock management systems: SAP EWM, barcode scanners and RFID·Project management: planning, milestone tracking, stakeholder updates and delivery·Rehabilitation and mobility support: exercises, aids and progress tracking· | |
| Sofia | Marchetti | Event Coordinator / Fitness Instructor / Travel Consultant | B | Spanish Native, Polish C2 | ·Stock management systems: SAP EWM, barcode scanners and RFID·Supply chain coordination: procurement, supplier management and delivery scheduling·Digital tools: Microsoft 365, scheduling software, CRM and ticketing systems·Data analysis: SQL, Python (pandas), Power BI and Tableau·Medication administration and dosage management·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Electrical installation: wiring, testing and fault diagnosis· | |
| Carlos | Mendez | Electrician / Solar Installer / Building Inspector | B | German Native, Arabic B2, Italian C1, French C2 | ·Administrative support: diary management, minute-taking and correspondence·Financial reporting: P&L, cash-flow analysis and month-end close·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Procurement and invoicing: purchase orders, reconciliation and vendor records·Attention to detail and accuracy in repetitive tasks·Data entry and records management: accuracy, GDPR-compliant handling·Project management: planning, milestone tracking, stakeholder updates and delivery· | |
| Samuel | Mensah | Bus Driver / Logistics Dispatcher / Tour Guide | A2, B, C1, C1E, D, DE | German Native, Spanish B1, Italian C2, French C1 | ·Automation: scripting, workflow tools and no-code integrations·Administrative support: diary management, minute-taking and correspondence·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Budget ownership: cost control, forecasting and variance reporting·Presentation and public speaking to technical and non-technical audiences·Security operations: access control, CCTV monitoring, patrols and incident response· | |
| Julia | Novak | Midwife / Lactation Consultant / Antenatal Educator | AM, B, C1 | Spanish Native, Polish C2 | ·Administrative support: diary management, minute-taking and correspondence·Problem solving: root-cause analysis and practical decision making·Presentation and public speaking to technical and non-technical audiences·Health & safety compliance: risk assessments, incident reporting, first aid trained·Facilities management: contractor coordination, planned and reactive maintenance·Data entry and records management: accuracy, GDPR-compliant handling·Content creation: copywriting, social media and photography· | |
| Tomasz | Nowak | Electrician / Solar PV Technician / Site Supervisor | A, B | English Native, German A2, Polish C1, Arabic A2 | ·Emergency response: evacuation procedures, fire safety and incident coordination·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Patient care: assessment, monitoring and daily living support·Data entry and records management: accuracy, GDPR-compliant handling·Attention to detail and accuracy in repetitive tasks·Administrative support: diary management, minute-taking and correspondence·Automation: scripting, workflow tools and no-code integrations· | |
| Clara | Nowicka | Nurse Assistant / Care Coordinator / Fitness Coach | B, D, DE | Spanish Native, English C2, Italian B2 | ·Supply chain coordination: procurement, supplier management and delivery scheduling·Security operations: access control, CCTV monitoring, patrols and incident response·Plumbing and HVAC basics: repairs, servicing and inspections·Financial reporting: P&L, cash-flow analysis and month-end close·Data entry and records management: accuracy, GDPR-compliant handling·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Administrative support: diary management, minute-taking and correspondence· | |
| David | Okonkwo | Accountant / Tax Advisor / Football Referee | A, B, C | Italian Native, French C2, English A2 | ·Technical writing: documentation, SOPs and knowledge-base articles·Supply chain coordination: procurement, supplier management and delivery scheduling·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Attention to detail and accuracy in repetitive tasks·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Project management: planning, milestone tracking, stakeholder updates and delivery·Web development: HTML, CSS, JavaScript and responsive design· | |
| Amara | Osei-Bonsu | Physiotherapist / Sports Rehabilitation Specialist / Health Coach | A, B | Spanish Native, German A2, English B1 | ·Administrative support: diary management, minute-taking and correspondence·Blueprint and technical drawing interpretation·Time management and prioritisation under deadline pressure·Plumbing and HVAC basics: repairs, servicing and inspections·Project management: planning, milestone tracking, stakeholder updates and delivery·Supply chain coordination: procurement, supplier management and delivery scheduling·Financial reporting: P&L, cash-flow analysis and month-end close· | |
| Viktor | Petrov | Chef de Partie / Baker / Food Photographer | B | French Native, Arabic C2, English B2 | ·Event management: budgeting, vendor sourcing, logistics and on-site delivery·Stakeholder communication: de-escalation, clear written and verbal updates, reporting·Problem solving: root-cause analysis and practical decision making·Cybersecurity awareness: access control, phishing prevention and data handling·Change management: rollout planning, training and adoption support·Emergency response: evacuation procedures, fire safety and incident coordination·Administrative support: diary management, minute-taking and correspondence· | |
| Priya | Ramesh | Data Analyst / Bookkeeper / Café Owner | B | Italian Native, English A1, Spanish A1, German A2 | ·Administrative support: diary management, minute-taking and correspondence·Emergency response: evacuation procedures, fire safety and incident coordination·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Manual handling and ergonomics: safe lifting and workstation assessment·Change management: rollout planning, training and adoption support·Blueprint and technical drawing interpretation·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control· | |
| Matteo | Rossi | Head Chef / Restaurant Manager / Sommelier | B | Arabic Native, French B1 | ·Health & safety compliance: risk assessments, incident reporting, first aid trained·Negotiation: supplier terms, contracts and conflict resolution·Forklift and reach-truck operation and instruction (Staplerschein trainer)·Stock management systems: SAP EWM, barcode scanners and RFID·Data entry and records management: accuracy, GDPR-compliant handling·Blueprint and technical drawing interpretation·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation· | |
| Rafael | Santos | Mechanic / Race Team Engineer / Driving Instructor | B, BE | Spanish Native, Italian C2 | ·Data analysis: SQL, Python (pandas), Power BI and Tableau·Electrical installation: wiring, testing and fault diagnosis·Performance management: goal setting, appraisals and constructive feedback·Project management: planning, milestone tracking, stakeholder updates and delivery·Problem solving: root-cause analysis and practical decision making·Technical writing: documentation, SOPs and knowledge-base articles·Change management: rollout planning, training and adoption support· | |
| Yuki | Sato | Lab Technician / Quality Analyst / Science Communicator | A, B, C | Arabic Native, Spanish C1, French B2, German A1 | ·Security operations: access control, CCTV monitoring, patrols and incident response·Cybersecurity awareness: access control, phishing prevention and data handling·Procurement and invoicing: purchase orders, reconciliation and vendor records·IT support: troubleshooting, hardware setup and user account administration·Data analysis: SQL, Python (pandas), Power BI and Tableau·Stakeholder communication: de-escalation, clear written and verbal updates, reporting·Content creation: copywriting, social media and photography· | |
| Birgit | Schulz | Delivery Driver / Cook / Cloud Architect | A1, B, BE, C | Spanish Native, English A2, Italian B1 | ·Health & safety compliance: risk assessments, incident reporting, first aid trained·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·UX/UI design: wireframing, prototyping and user research·Digital tools: Microsoft 365, scheduling software, CRM and ticketing systems·Supply chain coordination: procurement, supplier management and delivery scheduling·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Process improvement: workflow optimisation, cost reduction, KPI tracking and reporting·Adaptability: thriving in fast-changing and multi-tasking environments· | |
| Yuki | Tanaka | UX Designer / Frontend Developer / Creative Director | A2, B, C1 | French Native, German B2, English A2, Italian B2 | ·Bookkeeping and accounting: DATEV, Lexoffice, VAT/USt filings and payroll preparation·Presentation and public speaking to technical and non-technical audiences·Technical writing: documentation, SOPs and knowledge-base articles·Stock management systems: SAP EWM, barcode scanners and RFID·Facilities management: contractor coordination, planned and reactive maintenance·Cybersecurity awareness: access control, phishing prevention and data handling· | |
| Elena | Vasquez | Marketing Manager / Yoga Instructor / Bartender | B, C, CE | German Native, English B2 | ·Spreadsheets: Excel/Power Query, pivot tables, formulas and dashboards·Data analysis: SQL, Python (pandas), Power BI and Tableau·Warehouse operations: goods-in, slotting, picking, dispatch and inventory control·Web development: HTML, CSS, JavaScript and responsive design·Time management and prioritisation under deadline pressure·Cloud platforms: AWS, Azure and Google Cloud fundamentals·Digital tools: Microsoft 365, scheduling software, CRM and ticketing systems· | |
| Jonas | Weber | Software Developer / DJ / Personal Trainer | B | Spanish Native, French B2, Polish A1, German A2 | ·IT support: troubleshooting, hardware setup and user account administration·UX/UI design: wireframing, prototyping and user research·Facilities management: contractor coordination, planned and reactive maintenance·Presentation and public speaking to technical and non-technical audiences·Adaptability: thriving in fast-changing and multi-tasking environments·Administrative support: diary management, minute-taking and correspondence·Multilingual communication: fluent cross-language written and spoken support· |
In proRAG Query öffnen, den letzten Tab Both wählen, die Parameter zunächst unverändert lassen. Dieselbe Frage in lightRAG stellen, jede Antwort gegen die Galerie prüfen.
1) Wie viele Lebensläufe sind insgesamt im System gespeichert? 2) Welche Personen haben Führerschein BE UND sprechen English? Namen nennen. 3) Wie viele Personen haben KEINEN Führerschein C2? 4) Wie verteilen sich die Führerscheinklassen über alle Bewerber? 5) Welche Personen nennen sowohl Presentation and public speaking to technical and non-technical audiences als auch Administrative support als Skill? Namen nennen. 6) Welche Sprache wird am seltensten gesprochen, und von wie vielen Personen?
Frage 5 hat keine Auswahlliste: den Skill im Suchfeld der Galerie in doppelte Anführungszeichen setzen, sonst zerlegt die Suche ihn in Wörter und trifft zu viel.
| # | proRAG | lightRAG | Korrekt/Plausibel? | Qualitätsurteil |
|---|---|---|---|---|
| 1 | ___ | ___ | Tatsächlich: 40 | ___ |
| 2 | ___ | ___ | ___ | ___ |
| 3 | ___ | ___ | ___ | ___ |
| 4 | ___ | ___ | ___ | ___ |
| 5 | ___ | ___ | ___ | ___ |
| 6 | ___ | ___ | ___ | ___ |
Fragen aus Schritt 1 mit veränderten Parametern wiederholen. KI darf helfen, aber jede Behauptung muss gegen die Galerie prüfbar sein.
| Bedienelement in proRAG | Auswahl | Standard |
|---|---|---|
Search Mode | vector · bm25 · hybrid | hybrid |
Query Strategy | direct · hyde · multi_query · step_back | direct |
Temperature | — | 0.1 |
Retrieval Top K | — | 20 |
Reranker | — | ✅ |
Reranker Top N | — | 5 |
1) Worin unterscheiden sich die beiden angebotenen Suchmodi? 2) Was bewirken die verschiedenen Query Strategies? 3) Was verändert die Temperature am Ergebnis? Wann niedrig, wann hoch? 4) Was steuert Retrieval Top K, und welcher Fragetyp hängt massiv davon ab? 5) Was ist ein Reranker? Was ändert sich aktiv gegenüber inaktiv, und was passiert, wenn der Standardwert von Reranker Top N verändert wird?
1) In lightRAG die Graph-Ansicht öffnen. 2) Den Knoten English suchen. 3) Seine Kanten zu Personen zählen. 4) Stimmt die Zahl mit dem überein, was die Galerie-Filter in Schritt 1 zeigen?
So wurde der Korpus erzeugt:
Create new CV based on the architecture of the ones in the folder. Apply
random values from JSON files to fill up user skills, driving license and
language. Every CV can have one or more driving and languages and skills
can have from 5 to 8.
# languages.json
German, English, French, Spanish, Italian, Polish, Arabic
a) Warum hat dieser Prompt zusammen mit den JSON-Dateien Lebensläufe erzeugt, die für diese Übung optimal sind — und warum sind echte Lebensläufe höchstwahrscheinlich anders aufgebaut? b) Welche Massnahmen kann man in der Realität ergreifen, um echte Lebensläufe sauber in ein RAG einzubetten, ohne sich von KI-erzeugten Lebensläufen manipulieren zu lassen? Warum ist gerade der Embedding-Prozess bei unstrukturierten Daten so kritisch?
| # | Frage | Referenzantwort | RAG-typische Abweichung |
|---|---|---|---|
| 1 | Wie viele Lebensläufe insgesamt? | 40 | RAG antwortet typisch mit einem Bruchteil davon (Top-K-Limitierung — sieht nie alle Dokumente) |
| 2 | Führerschein BE UND English? | 6 Personen: Kwame Adeyemi, Henrik Andersen, Marco Fernandez, Anna Kowalczyk, Mei Ling Chen, Birgit Schulz | Der Mehrattribut-Filter über Dokumente hinweg ist die klassische RAG-Schwäche: meist fehlen Personen, gelegentlich kommt eine dazu, die nur eines der beiden Kriterien erfüllt. |
| 3 | Kein Führerschein C2? | Diese Führerscheinklasse existiert nicht — richtig wäre: nicht vorhanden, unbekannte Klasse. | Beide Systeme antworten meist mit einer Zahl. Ähnlichkeitssuche kennt kein "gibt es nicht": die Frage sieht aus wie die anderen, also wird sie wie die anderen beantwortet. |
| 4 | Verteilung der Führerscheinklassen? | AM 3 · A1 5 · A2 3 · A 6 · B 40 · BE 9 · C1 8 · C1E 4 · C 6 · CE 1 · D1 1 · D 7 · DE 5 | Eine Verteilung setzt voraus, dass ALLE Dokumente gesehen wurden — genau das leistet Top-K nicht. Die seltenen Klassen fehlen fast immer. |
| 5 | Presentation and public speaking to technical and non-technical audiences UND Administrative support? | 4 Personen: Leila Hassan, Samuel Mensah, Julia Novak, Jonas Weber | Ko-Vorkommen zweier Skills: RAG retrievt zu jedem Skill die ähnlichsten Chunks und schneidet sie schlecht — es nennt oft Personen mit nur einem der beiden. |
| 6 | Seltenste Sprache? | Polish & Arabic — je 13 Personen | Der Rand der Verteilung ist genau das, was Relevanz-Ranking wegschneidet. Achtung: hier gibt es einen Gleichstand — eine einzelne Sprache ist eine unvollständige Antwort. |
Die Beschreibungen aus der proRAG-Oberfläche, wörtlich:
Search Mode
vector Semantic similarity search using OpenAI embeddings. Best for
meaning-based queries.
bm25 Keyword-based search using BM25 algorithm. Best for exact
term matching.
hybrid Combines vector and BM25 results using Reciprocal Rank
Fusion. Best overall.
Query Strategy
direct Send the query directly to the retrieval pipeline without
transformation.
hyde Generate a Hypothetical Document Embedding — the LLM writes
a hypothetical answer, then uses its embedding for retrieval.
multi_query Generate multiple query variations to capture different
perspectives and improve recall.
step_back Rephrase the query at a higher abstraction level to retrieve
broader context.
cv_manifest.json abgeleitet, nicht von Hand gepflegt: die gefragte Führerscheinklasse, die Sprache und die beiden Skills sind so gewählt, dass sie den Korpus wirklich teilen — eine Frage, die auf alle 40 zutrifft, prüft nichts. Wird der Korpus getauscht, wählt die Seite neu. Jede Antwort ist mit den Galerie-Filtern in Schritt 0 nachprüfbar. Einzige Ausnahme ist Frage 3: die dort gefragte Klasse ist bewusst festgeschrieben, weil eine abgeleitete Klasse irgendwann eine echte wäre.
Diese Beobachtung stammt aus einer früheren, deutlich grösseren Fassung dieser Aufgabe: Zwei unabhängige KI-Instanzen (Claude Opus) erstellten die damaligen Referenzantworten. Eine las die Lebensläufe in kleinen Gruppen sorgfältig durch, die andere (Codex-Verifier) versuchte den gesamten Korpus auf einmal zu analysieren. Ergebnis: Der Codex-Verifier halluzinierte in mehreren Fällen — er behauptete z.B., Personen hätten SQL-Kenntnisse, obwohl deren CV kein SQL erwähnt, oder zählte Führerscheine bei Personen, die gar kein Führerschein-Feld im CV hatten. Die sorgfältigere Analyse (kleinere Batches, mehr Aufmerksamkeit pro CV) war in allen strittigen Fällen korrekt.
Lektion für die Praxis: Das gleiche Problem, das RAG-Systeme bei diesen Fragen haben — unvollständige Erfassung, falsche Zuordnung, Halluzination bei Aggregation — betrifft auch grosse Sprachmodelle direkt. Je mehr Dokumente auf einmal verarbeitet werden, desto höher das Halluzinations-Risiko. Teile-und-Herrsche (Dokumente in Gruppen aufteilen) liefert zuverlässigere Ergebnisse als alles auf einmal zu verarbeiten.
Bisher haben Sie nur GESUCHT — die Einbettung (Embedding) neuer Dokumente blieb eine Blackbox. Diese Zusatzaufgabe schließt den Kreis: Sie erzeugen einen neuen Lebenslauf, laden ihn hoch und prüfen per Abfrage, dass er wirklich indexiert wurde.
Öffnen Sie Claude Code im WSL-Terminal (claude) — es ist auf den Lab-Rechnern vorinstalliert — und fügen Sie diesen Prompt ein:
Create a job-application CV as a PDF for a FICTIONAL person applying to a pizzeria as a delivery driver OR a cook (pick one). Age between 20 and 40. Include a realistic, role-appropriate CV: personal data, work experience, skills — driving licence and vehicle for a driver, kitchen skills for a cook. Write a small Python script to produce the PDF and name the file Firstname_Lastname_Role.pdf (ASCII only, underscores instead of spaces). Show me the full path of the generated PDF.
Melden Sie sich an proRAG an (Zugangsdaten: siehe Login-Abschnitt oben), laden Sie das PDF über den Dokument-Upload der Weboberfläche hoch und warten Sie, bis die Verarbeitung (Parsing + Embedding) abgeschlossen ist.
Stolperstein: Das PDF liegt IN WSL2 (Ubuntu), Ihr Browser läuft aber unter Windows. Öffnen Sie den WSL-Ordner direkt im Windows Explorer — oder tippen Sie den WSL-Pfad in die Adressleiste des Explorers bzw. des Datei-Dialogs und wählen Sie das PDF dort im Upload-Dialog aus:
# In the WSL directory holding the PDF — opens Windows Explorer right there explorer.exe . # Or type the WSL path into the Explorer / file-picker address bar # (older Windows builds: \\wsl$\... instead of \\wsl.localhost\...) \\wsl.localhost\<distro>\home\<user>\Firstname_Lastname_Role.pdf
Öffnen Sie zum Abschluss den Chunking Playground und laden Sie dort Ihr EIGENES erzeugtes Lebenslauf-PDF hoch: Spielen Sie mit den verschiedenen Chunking-Strategien, aktivieren Sie den Comparison Mode und beobachten Sie, wie unterschiedlich Ihr Dokument je nach Strategie zerlegt (gechunkt) und eingebettet wird. Es gibt kein festes erwartetes Ergebnis — einfach ein bisschen rumspielen und schauen, was die Strategien mit Ihrem Dokument machen.