Home
← Internships

Agentic GraphRAG with Adaptive Graph Traversal for EPC Tendering

Technip Energies · Mar 2026 – Jul 2026 · Supervised by Julien Cadart

Defense slides (PDF)

Context

Tendering for EPC (engineering, procurement, and construction) projects produces a large, multidisciplinary corpus: engineering, procurement, project management, legal, finance, insurance, and safety documents spanning about 40 specialised domains — from rotating equipment and piping to document control and risk management — in heterogeneous formats (PDF, Word, Excel, PowerPoint…).

Finding information across this corpus is manual, cross-checking is time-consuming, and critical information is easy to miss. The internship asked a concrete research question: can knowledge-graph-based retrieval improve the accuracy and completeness of the existing RAG system for large-scale engineering tender documents?

Contributions

  • Domain-aware knowledge graph construction from a heterogeneous tender corpus
  • An adaptive graph retrieval pipeline that picks a strategy based on the query
  • Dual-Level RAG and Weighted Hub-aware PPR-RAG, combined with a cross-encoder reranker
  • Quantitative and qualitative evaluation against a Hybrid RAG baseline

How the system works

The pipeline builds a domain-aware knowledge graph from the tender corpus (free-form semantic relation extraction, constrained to entity types like deliverable, risk, requirement, project, and system), then answers queries in four stages:

  1. Query understanding. Each query is classified by retrieval need (textual factoid, single-hop lookup, aggregation/summarization, or relation discovery/multi-hop) and mapped to a graph scope: single domain, multi-domain, or full graph.
  2. Entity linking. Entities extracted from the query are matched to graph nodes with fuzzy matching and semantic search, fused via reciprocal rank fusion, then disambiguated and deduplicated by an LLM to produce seed entities.
  3. Adaptive retrieval. Based on the query type, the system routes to Hub-aware PPR retrieval, local graph retrieval, Dual-Level retrieval, or plain Hybrid RAG (BM25 + dense vectors on Azure AI Search) — avoiding expensive graph traversal when it isn't needed.
  4. Fusion and answering. Graph evidence and hybrid-search evidence are fused and deduplicated, reranked by a cross-encoder, and passed to an LLM for answer generation.
System architecture: a query flows through query understanding, domain prediction, and retrieval-need classification into one of four retrieval strategies, whose evidence is fused, reranked by a cross-encoder, and answered by an LLM.
Overall system architecture: adaptive routing between graph-based retrieval strategies and Hybrid RAG.

Results

Evaluation used 200 questions, an LLM-as-a-judge protocol, and a Hybrid RAG baseline (BM25 + dense vectors). Dual-Level RAG clearly improved summarisation and aggregation queries:

Hybrid RAGDual-Level RAG
00.250.500.751Hybrid RAG — Accuracy: 0.840.840Dual-Level RAG — Accuracy: 0.9650.965AccuracyHybrid RAG — Completeness: 0.7940.794Dual-Level RAG — Completeness: 0.8820.882CompletenessHybrid RAG — Relevance: 0.930.930Dual-Level RAG — Relevance: 0.9880.988RelevanceHybrid RAG — Context precision: 0.8240.824Dual-Level RAG — Context precision: 0.8570.857Context precision
MetricHybrid RAGDual-Level RAG
Accuracy0.8400.965
Completeness0.7940.882
Relevance0.9300.988
Context precision0.8240.857
Latency20 s31 s
Dual-Level RAG vs the Hybrid RAG baseline, across all 200 questions.

For relational and multi-hop queries, Weighted Hub-aware PPR-RAG brought smaller, not clearly significant gains — at almost double the latency:

Hybrid RAGPPR-RAG
00.250.500.751Hybrid RAG — Accuracy: 0.8750.875PPR-RAG — Accuracy: 0.8820.882AccuracyHybrid RAG — Completeness: 0.850.850PPR-RAG — Completeness: 0.8620.862CompletenessHybrid RAG — Relevance: 0.9050.905PPR-RAG — Relevance: 0.9510.951RelevanceHybrid RAG — Context precision: 0.8050.805PPR-RAG — Context precision: 0.8410.841Context precision
MetricHybrid RAGPPR-RAG
Accuracy0.8750.882
Completeness0.8500.862
Relevance0.9050.951
Context precision0.8050.841
Latency21 s37 s
Weighted Hub-aware PPR-RAG vs Hybrid RAG, on relational and multi-hop queries only.

Limitations

Failure analysis pointed to the graph itself: the constructed knowledge graph had 3,921 connected components, and 75.2% of them contained five nodes or fewer. Most entities were isolated or lived in tiny neighbourhoods, so Personalized PageRank could not propagate across disconnected components. Imperfect entity resolution (acronyms, aliases, inconsistent terminology) was a major cause of that fragmentation.

Takeaway

Knowledge graphs improve information aggregation over plain Hybrid RAG, and graph retrieval and hybrid retrieval are complementary rather than competing. But reliable multi-hop retrieval requires a cleaner, better-connected graph — pointing to future work on ontology refinement, entity resolution (acronym and alias normalisation, merging semantically equivalent entities), a gold-standard question–answer benchmark, and evaluation on other EPC tenders.