1. Challenge Summary
The challenge aims to design, implement and evaluate a question/answering (QA) system running exclusively on CPU, based on a purely symbolic architecture (lexicon, grammars, knowledge graph, formal logic). The entire system development is driven by quantised open-weights LLMs, executable on consumer hardware (i9 CPU, RTX 3090).
This project falls within the Green AI movement (Schwartz et al., 2020) and Frugal AI, in opposition to the race towards ever larger models whose carbon footprint has become alarming (Kim et al., 2026; Hasan et al., 2025). It also addresses the growing demands for transparency, explainability and regulatory compliance imposed by the GDPR and the European AI Act (Wagner et al., 2025; Juliussen, 2025).
- Purely symbolic QA system (no neural network at inference)
- CPU-only at runtime (no GPU required)
- Automatic KG construction from natural language documents
- Development entirely supervised by quantised open-weights LLMs
- Native GDPR/AI Act compliance by design (intrinsic interpretability)
- Ethical license (Hippocratic 3.0)
2. Context and Scientific Positioning
2.1. Problem Statement
Current QA systems rely heavily on LLMs (GPT, LLaMA, etc.) that require expensive GPUs for inference, and raise issues of hallucination, reproducibility, explainability and data sovereignty. They operate as black boxes and are difficult to make compliant with the requirements of the European AI Act (Article 13) regarding the transparency of high-risk systems (Wagner, 2025; Boshnjakovski, 2025).
2.2. State of the Art
Neuro-symbolic Approaches and KG-QA
Recent literature converges towards integrating knowledge graphs (KG) and symbolic reasoning to address the limitations of purely neural approaches:
- Cheng et al. (2025), "Neural-symbolic methods for knowledge graph reasoning" (ACM TKDD) — distinguish KG completion, complex query answering and logical rule learning approaches; identify the potential of pure symbolic methods for interpretability.
- Niu et al. (2026), "A Comprehensive Survey of Knowledge Graph Reasoning" (IEEE Trans. Big Data) — cover multi-hop reasoning on KG, including Inductive Logic Programming (ILP) approaches.
- Linders & Tomczak (2025), "Knowledge graph-extended RAG for QA" (Applied Intelligence) — demonstrate that KGs enable symbolic reasoning and verifiable answers.
- Chudasama et al. (2025), "Toward interpretable hybrid AI" (IEEE Access) — propose TrustKG combining symbolic interpretability with KG.
- Agarwal & Bedathur (2025), "A Zero-Shot Neuro-Symbolic Approach for Complex KGQA" (EMNLP Findings 2025).
- Liu et al. (2025), "SymAgent" (TheWebConf 2025) — self-learning symbolic agent for KGQA.
Automatic Knowledge Graph Construction
- Zavarella (2026), "Methods for KG Construction from Text Collections" (arXiv) — framework combining NLP, ML and Semantic Web.
- Salman et al. (2024), "Doc-KG" (Expert Systems) — KG construction from unstructured documents.
- Chen et al. (2025), "Open Local Knowledge Graph Construction from Academic Papers Using Generative LLMs" (TheWebConf 2025).
CPU Efficiency and Lightweight Approaches
- Sidhu et al. (2025), "Neurosymbolic AI for Enhanced Reasoning in Small Language Models" (IEEE).
- Corradini et al. (2025), "State of the art and future directions of small language models" (Big Data and Cognitive Computing).
- Chaudhary et al. (2025) confirm that "spaCy provides lightweight, high-precision extraction, efficient on standard CPU".
- Okechukwu et al. (2025) confirm that "spaCy and NLTK operate efficiently on standard CPUs".
Ecological Impact of AI
- Kim & Yoo (2026), "Toward Sustainable AI" (IEEE Access) — the carbon footprint of AI data centres tripled between 2024 and 2025.
- Hasan et al. (2025), "Carbon emission quantification of ML" (IEEE Trans.).
- Tunzina et al. (2026), "GreenAI: Environmental Efficiency in LLM-Generated Code" (IEEE Access).
GDPR and AI Act Compliance
- Wagner et al. (2025), "A Critical Look at the EU AI Act's Requirements" (IEEE RE).
- Juliussen (2025), "The Right to an Explanation Under the GDPR and the AI Act" (Springer).
- Goncalves & Correia (2026), "XAI-Compliance-by-Design" (MDPI).
- Sovrano et al. (2025), "Legal XAI: Systematic Review" (ResearchGate).
3. Objectives
Main Objective
Demonstrate the feasibility and evaluate the performance of a CPU-only, purely symbolic QA system whose knowledge graph is built automatically from natural language documents.
Secondary Objectives
- Achieve competitive performance on recognised KGQA benchmarks.
- Measure and compare the energy footprint (CPU vs GPU) to demonstrate the ecological benefit.
- Demonstrate the intrinsic interpretability of the system (traceability of each answer).
- Provide reproducible, documented software under an ethical license.
4. Architecture & Technical Specifications
4.1. Overall Architecture
The system is divided into two distinct phases:
Phase 1 — Knowledge Graph Construction (offline, GPU allowed)
NL Documents → Preprocessing → NER → Relation Extraction → Coreference Resolution
→ KG Construction (RDF/OWL triples) → Graph Storage
Phase 2 — Question/Answering (runtime, strict CPU-only)
NL Question → Syntactic/Semantic Analysis → Transformation to Formal Query
→ KG Traversal → Logical Reasoning → Answer Generation
→ Explanation Trace
4.2. Hardware Specifications
| Component | Phase 1 (KG Construction) | Phase 2 (QA Runtime) |
|---|---|---|
| CPU | Intel Core i9 (max) | Intel Core i9 (max) |
| GPU | NVIDIA RTX 3090 (24 GB VRAM) | None (strict CPU-only) |
| RAM | 32 GB (max) | 32 GB (sufficient) |
| Storage | NVMe SSD | NVMe SSD |
Justification: Phase 1 may use the GPU to run the quantised open-weights LLMs that supervise pipeline construction. Phase 2, which constitutes the deployable system, must be strictly CPU-only to guarantee accessibility, sovereignty and frugality.
4.3. Development Environment
- Sandbox: development in an isolated environment for security and reproducibility.
- Target OS: Linux (primary target), with Windows/macOS portability desired.
- Language(s): free choice, provided dependencies comply with license constraints (section 6).
4.4. LLM Supervision
System development is entirely driven by one or more open-weights LLMs, executable locally on the specified hardware (Phase 1). The choice of models, quantisation and inference framework is left to the participant.
5. Suggested Pipeline Architecture
The following breakdown is provided as a guideline. Participants are free to propose a different architecture, provided they meet the functional constraints (sections 3, 4 and 12) and license requirements (section 6).
5.1. Text Preprocessing
- Sentence segmentation
- Tokenisation
- Lemmatisation
- Part-of-Speech tagging
- Dependency parsing
5.2. Named Entity Recognition (NER)
- Entity detection (persons, locations, organisations, dates, concepts)
- Entity typing according to a domain ontology
- Coreference resolution
5.3. Relation Extraction (RE)
- Triple extraction (subject, predicate, object) from text
- Open or pattern-guided extraction
5.4. Knowledge Graph Construction
- Entity normalisation and merging (entity linking)
- Graph construction and storage (triples)
- Enrichment through implicit relation inference
- Graph validation and cleaning
5.5. Question Understanding
- Syntactic analysis of the natural language question
- Question type identification (who, what, where, when…)
- Extraction of mentioned entities and relations
- Transformation into a formal query on the graph
5.6. Reasoning & Answer Search
- Knowledge graph traversal (multi-hop reasoning)
- Logical resolution and ranking of candidate answers
5.7. Answer Generation & Explanations
- Answer formulation in natural language
- Complete explanation trace: path in the KG, applied rules, involved entities and relations, confidence/completeness
6. License Constraints
6.1. Hippocratic License 3.0 (HL3)
The final software will be published under Hippocratic License 3.0 (firstdonoharm.dev). This license:
- Is not considered open source in the OSI sense (ethical constraints on usage).
- Is a copyleft license (Section 8.3).
- Imposes ethical standards based on fundamental human rights (UDHR, ICESCR, ICCPR, ILO).
6.2. Dependency Compatibility
| License | HL3 Compatible | Examples |
|---|---|---|
| MIT | ✅ Yes | — |
| BSD (2/3-clause) | ✅ Yes | — |
| Apache 2.0 | ✅ Yes | — |
| ISC / CC0 | ✅ Yes | — |
| GPL v2/v3 | ⚠️ Potentially incompatible | — |
| AGPL | ❌ Probably incompatible | — |
6.3. Recommended Optional HL3 Modules
- HL3-ECO (Ecocide) — consistent with the objective of reducing ecological impact.
- HL3-MEDIA — prohibition of use for promoting violence.
- HL3-SV (Mass Surveillance) — consistent with sovereignty and privacy.
7. Regulatory Compliance
7.1. GDPR
| GDPR Principle | Implementation |
|---|---|
| Data minimisation (Art. 5.1.c) | The KG stores only relevant extracted triples, not raw personal data. |
| Right to explanation (Art. 22 + recital 71) | Complete explanation trace for each answer. |
| Right to erasure (Art. 17) | Ability to delete entities/triples from the KG. |
| Privacy by design (Art. 25) | No sensitive data in the KG unless documented necessity. |
| No automated profiling (Art. 22) | The system does not make automated decisions about individuals. |
7.2. European AI Act
| AI Act Requirement | Compliance |
|---|---|
| Transparency (Art. 13) | Intrinsically interpretable symbolic architecture — every reasoning step is traceable. |
| Explainability (Art. 13.1) | Complete trace: question → formal query → KG path → answer. |
| Technical documentation (Art. 11) | Fully documentable symbolic pipeline (grammars, rules, ontology). |
| Robustness and accuracy (Art. 15) | No hallucination — answers come exclusively from the KG. |
| Human oversight (Art. 14) | The KG is inspectable and modifiable by a human. |
| Risk management (Art. 9) | Low risk: no opaque model, no emergent bias. |
8. Ecological Impact
8.1. The Problem
- Training GPT-3 consumed approximately 1,287 MWh, i.e. ~552 tonnes of CO₂.
- AI data centres saw their carbon footprint triple between 2024 and 2025 (Awab, 2026).
- An LLM query consumes approximately 10× more energy than a standard web search (Kim & Yoo, 2026).
8.2. Advantage of Symbolic CPU-only
| Metric | LLM System (GPU) | AI Challenge (CPU-only) | Estimated Gain |
|---|---|---|---|
| Instantaneous Power | 300-400 W (GPU) | 65-125 W (CPU) | ~3-5× |
| Latency per Query | 0.5-5 s | < 100 ms (target) | ~5-50× |
| Energy per Query | ~0.001-0.01 kWh | ~0.000001-0.00001 kWh | ~100-1000× |
| Required Hardware | GPU ≥ €6,000 | Consumer CPU < €500 | ~12× |
8.3. Mandatory Measurement
The project must measure and report:
- Energy consumption (kWh) for KG construction.
- Energy consumption per QA query.
- Corresponding CO₂ footprint.
- Comparison with a GPU baseline system.
9. Reproducibility
- Complete source code: published on a public repository, under HL3 license.
- Installation instructions: automated script or containerisation.
- Test data: exclusive use of public benchmarks.
- Random seeds: fixed and documented for any stochastic component.
- Dependency versioning: pinned dependency file.
- Containerisation: the complete system must be reproducibly executable.
- Documented minimum hardware: explicit minimum specifications.
The maximum required hardware (i9 + RTX 3090) is accessible to a wide audience. Phase 2 (QA runtime) requires no GPU, enabling execution on a simple laptop.
10. Sovereignty, Security & Privacy
Digital Sovereignty
- No cloud dependency — fully local operation.
- No external API calls — no data leaves the machine.
- Open-weights LLMs downloadable and executable locally.
- No vendor lock-in.
Security
- Sandboxed execution (isolated environment).
- No arbitrary code execution.
- No prompt injection (no LLM at inference).
- Input validation via formal grammar.
Privacy
- Exclusive local processing — documents never leave the machine.
- No telemetry.
- KG erasable on demand.
- Ability to instantiate separate KGs per user or per domain.
11. Benchmarks & Evaluation
11.1. Recognised KGQA Benchmarks
The system must be evaluated on at least two of the following benchmarks:
| Benchmark | Description | Size | Reference |
|---|---|---|---|
| SimpleQuestions | Simple factual questions on Freebase | ~100K questions | ACL |
| WebQuestionsSP | Web questions with SPARQL annotations | ~4,737 questions | ACL |
| LC-QuAD 2.0 | Complex questions on DBpedia/Wikidata | ~30K questions | ISWC |
| ComplexWebQuestions | Multi-hop questions on Freebase | ~34K questions | NAACL |
| QALD | Multilingual challenge, small size | ~400-500 questions | CLEF/ESWC |
| MINTQA | Recent benchmark for LLM+KG on rare entities | Variable | EMNLP 2025 |
| SQuAD 2.0 | Extractive QA on paragraphs | ~150K questions | ACL |
| Natural Questions | Real Google questions | ~307K questions | TACL |
| TriviaQA | Trivia QA with supporting documents | ~95K questions | ACL |
Recommendation: prioritise LC-QuAD 2.0 and QALD, supplemented by SQuAD or Natural Questions for comparison with neural approaches.
11.2. Evaluation Metrics
| Category | Metric | Description |
|---|---|---|
| QA Performance | F1 score, Exact Match (EM) | Answer precision |
| QA Performance | Hits@1, Hits@5 | Correct answer rate in top-k |
| QA Performance | MRR | Mean rank of correct answer |
| Latency | Average/p95 response time (ms) | Measured on CPU only |
| Throughput | Questions/second | Throughput |
| Ecology | kWh/query, gCO₂/query | Via codecarbon/carbontracker |
| Explainability | Traceability rate | % of answers with complete trace (target: 100%) |
| KG Coverage | Triples, entities, relations | Descriptive statistics |
11.3. Comparison Baselines
- RAG + LLM (GPU): neural baseline on GPU.
- KG-RAG (GPU): hybrid KG + LLM baseline.
- Direct SPARQL: pure symbolic baseline (manual queries on the same KG).
12. Performance Specifications
| Criterion | Target | Justification |
|---|---|---|
| QA Latency | < 100 ms (p95) | Acceptable user interactivity |
| KG Construction | < 1 min / 1,000 short documents | Practicality |
| QA Memory Footprint | < 4 GB RAM | Executable on modest PC |
| KG Size | Scalable to 10M+ triples | Sufficient coverage for benchmarks |
| QA Precision | ≥ 50% F1 on simple benchmarks | Interpretability and frugality as main arguments |
| Traceability | 100% of answers explained | Key advantage of symbolic approach |
13. Risks & Mitigations
| Risk | Probability | Impact | Mitigation |
|---|---|---|---|
| Insufficient QA performance vs LLM | High | Medium | Focus on interpretability/frugality |
| Insufficient quality of automatic KG | Medium | High | Iterate rules, LLM supervision |
| Multi-hop question complexity | High | Medium | Limit to ≤ 3 hops, document limitations |
| Dependency license incompatibility | Low | High | Systematic license audit |
| Open-weights LLM insufficient for supervision | Medium | Medium | Test and combine multiple models |
| Limited KG coverage at scale | Medium | Medium | Start with QALD, then scale up |
14. Success Criteria
- The QA system operates entirely on CPU with no GPU at inference.
- The KG is built automatically from natural language documents.
- Performance is measured on ≥ 2 recognised public benchmarks.
- 100% of answers are accompanied by an explanation trace.
- Energy consumption is measured and compared to a GPU baseline.
- Code is published under Hippocratic 3.0 license with compatible dependencies.
- The system is reproducible via Docker on the specified hardware.
- Development is entirely carried out by open-weights LLMs in a sandbox.
15. Bibliographic References
- Agarwal, P. & Bedathur, S. (2025). A Zero-Shot Neuro-Symbolic Approach for Complex Knowledge Graph Question Answering. EMNLP 2025 Findings.
- Ansar, W., Goswami, S. & Chakrabarti, A. (2024). From Transformers to LLMs: A Systematic Survey of Efficiency Considerations in NLP. arXiv:2406.16893.
- Awab, S. (2026). AI's Resource Gluttony and Sustainable Development. SSRN 6301563.
- Boshnjakovski, V. (2025). Balancing Compliance and Innovation: Addressing Explainability in AI Systems under the EU AI Act. MBA Thesis.
- Chaudhary, V. et al. (2025). Fusing Knowledge and Language: A Comparative Study of KG-Based QA with LLMs. arXiv:2509.09272.
- Chen, H. et al. (2025). Open Local Knowledge Graph Construction from Academic Papers Using Generative LLMs. TheWebConf 2025.
- Cheng, K. et al. (2025). Neural-symbolic methods for knowledge graph reasoning: A survey. ACM Trans. Knowl. Discov. Data.
- Chudasama, Y. et al. (2025). Toward interpretable hybrid AI: integrating knowledge graphs and symbolic reasoning. IEEE Access.
- Corradini, F. et al. (2025). State of the art and future directions of small language models. Big Data and Cognitive Computing.
- Goncalves, A. & Correia, A. (2026). XAI-Compliance-by-Design: A Modular Framework for GDPR- and AI Act-Aligned Decision Transparency. MDPI J. Cybersecurity and Privacy.
- Hasan, S.M. et al. (2025). Carbon emission quantification of ML: A review. IEEE Trans. Sustainable Computing.
- Juliussen, B.A. (2025). The Right to an Explanation Under the GDPR and the AI Act. Springer, Int. Conf. Multimedia Modeling.
- Kim, M.K. & Yoo, T.A. (2026). Toward Sustainable AI: A Scoping Review of Carbon Footprint. IEEE Access.
- Linders, J. & Tomczak, J.M. (2025). Knowledge graph-extended RAG for QA. Applied Intelligence.
- Liu, B. et al. (2025). SymAgent: A neural-symbolic self-learning agent framework for complex reasoning over KGs. TheWebConf 2025.
- Ma, C. et al. (2025). Large language models meet knowledge graphs for question answering. EMNLP 2025.
- Niu, G. et al. (2026). A Comprehensive Survey of Knowledge Graph Reasoning. IEEE Trans. Big Data.
- Perevalov, A. et al. (2024). Multilingual question answering systems for knowledge graphs — a survey. Semantic Web.
- Salman, M. et al. (2024). Doc-KG: Unstructured documents to KG construction. Expert Systems.
- Schwartz, R. et al. (2020). Green AI. Communications of the ACM, 63(12), 54-63.
- Sidhu, M.K. et al. (2025). Neurosymbolic AI for Enhanced Reasoning in Small Language Models. IEEE.
- Sovrano, F. et al. (2025). Legal XAI: Systematic Review and Interdisciplinary Mapping of XAI and EU Law. ResearchGate.
- Tunzina, T. et al. (2026). GreenAI: Environmental Efficiency in LLM-Generated Code. IEEE Access.
- Wagner, M. et al. (2025). A Critical Look at the EU AI Act's Requirements — Who Must Explain What? IEEE Int. Requirements Eng. Conf.
- Zavarella, V. (2026). Methods for KG Construction from Text Collections. arXiv:2603.25862.