RAG
A RAG chunking strategy for technical manuals
Splitting a manual into arbitrary-length pieces makes the right answer get searched in the wrong context. Preserve headings, tables and revision information.
Making LLMs Reliable for Industrial Q&A: Evidence-First Design
General chatbots hallucinate; industrial answers must be verifiable. This article describes the evidence-first design pattern: curated corpora, hybrid retrieval, constrained synthesis, and inspectable output.
Why general chatbots fail in industry
A general chatbot answers from pattern-matching over its training data. In industrial contexts that is a liability: a confidently wrong answer about a PLC register map, a safety function, or a protocol's timeout behaviour is worse than no answer, because engineers may act on it. The failure mode is not the model's intelligence — it is the absence of verifiable grounding.
The evidence-first pattern
The reliable pattern for industrial Q&A is evidence-first:
- Curate a corpus: maintain versioned, domain-scoped datasets — process control, PLC/DCS, protocols, OPC, databases, ML — curated by engineers rather than scraped wholesale.
- Retrieve with hybrid ranking: combine keyword scoring (exact terms, identifiers, ports) with vector similarity (synonyms, concepts) to find the most relevant evidence.
- Constrain synthesis: the LLM receives the evidence packet and is instructed to answer from it, citing which source contributed, and to explicitly separate dataset-supported claims from general knowledge.
- Inspect output: every answer exposes its evidence — which sources were used, which tools ran, what the retrieval confidence was.
This is the architecture behind Kutup NQ: the browser runs intent classification and embedding locally, the backend merges dataset results with selective Wikipedia and RSS enrichment, and a synthesis layer formats the answer with visible evidence.
Designing for inspectability
The most important UX decision is making the pipeline visible. Engineers trust systems they can audit:
- Tool-use chips show which sources contributed (custom datasets, Wikipedia, RSS).
- Evidence panels expose the retrieved references, expandable to full text.
- Runtime details report retrieval confidence, model synthesis status, and rate-limit state.
- Fallback honesty — when the LLM is unavailable, the system says so and returns a dataset-sourced answer instead of pretending.
Inspectability is also a debugging tool: when an answer is wrong, the evidence panel shows exactly which retrieval step produced the bad context.
Bilingual and domain-scoped by design
Two constraints make industrial assistants practical:
- Bilingual by design: Turkish and English are first-class — language detection, dataset selection, and retrieval weighting all operate per language, including cross-language expansion.
- Domain-scoped boundaries: the assistant rejects out-of-scope prompts and only answers within its supported domains, trading breadth for reliability.
Key point: In industrial Q&A, the model is the least important component. The corpus, the retrieval, the constraints, and the evidence display determine whether the answer can be trusted — and whether mistakes are catchable.
Keep reading.
Full library
RAG
Industrial RAG: how to measure answer quality
Evaluate answers from engineering manuals with source relevance, device context and appropriate abstention.
RAG
Word Embeddings and Semantic Search: How Modern Retrieval Works
Semantic search matches meaning, not spelling. This article explains embeddings, vector similarity, and why hybrid keyword-plus-vector retrieval is the robust production pattern.