Automotive AI
Two million documents, retrieved with their structure intact.
I redesigned retrieval across two million ADAC documents, separating tables, legal guidance, vehicle tests, and editorial content into strategies that raised benchmark precision from 61% to 84%.
Technology
Two million indexed documents were not the same as usable knowledge
ADAC's corpus combined structured vehicle tests, dense legal guidance, comparison tables, and short editorial articles. A uniform retrieval strategy treated fundamentally different information as if it had the same shape.
The product needed trustworthy conversational answers with source attribution, strong German automotive terminology, and an audio path suitable for hands-free use.
Rebuilt retrieval around document structure
I owned work across corpus classification, chunking and retrieval strategy, reranking, answer guardrails, and the text-to-speech delivery path.
The initial benchmark made the architectural mistake visible: two million indexed documents were not useful when the relevant passage structure had been destroyed during chunking.
Architecture
Retrieval quality began with corpus architecture
Content type determined chunking, candidate generation, reranking, and answer constraints instead of forcing every source through one generic RAG pipeline.
- 01
Constraint
Uniform 512-token chunks split tables, legal clauses, and vehicle-test sections at arbitrary boundaries.
Decision
Classify documents by content type and apply table-aware, sentence-window, or section-aware chunking to each class.
Impact
Retrieval precision increased from 61% to 84% on the benchmark set.
- 02
Constraint
Approximately 8% of documents had missing or incorrect content-type metadata.
Decision
Combine a structural rules pass with a low-cost language-model classifier only for low-confidence documents.
Impact
The correct chunking strategy was applied to 98.7% of the corpus without using a model on every record.
- 03
Constraint
Dense retrieval alone missed exact model codes, tyre dimensions, and jurisdiction-specific terms.
Decision
Fuse vector and keyword retrieval, then rerank the combined candidates before generation.
Impact
Answers preserved semantic recall while improving exact-match behaviour and source relevance.
Outcome
Higher precision without brute-forcing every document through a model
- Benchmark retrieval precision reached 84%, up from a 61% baseline.
- The pilot resolved 73% of member queries without human escalation.
- Source attribution and guardrails kept answers tied to approved ADAC material.
- Automotive text-to-speech met the target intelligibility threshold with a maintained terminology dictionary.
RAG quality is often a data-modelling problem disguised as a model problem. The biggest improvement came from respecting the structure of the source material before retrieval began.
Abrar
Start a project
Need retrieval that works beyond a polished demo?
I can help benchmark the corpus, design the retrieval architecture, and build answer controls around the evidence your users actually need.
Discuss your RAG system