Pipeline Calibration

System Settings & RAG Tuning

Tune vector retrieval weights, neural reranker thresholds, and safety guardrails.

Retrieval & Hybrid Search

Parameters controlling candidate retrieval from the cosine vector space and subsequent cross-encoder precision filtering.

Initial Vector Candidates (Top-K)15 chunks

Raw candidate chunks retrieved from ChromaDB before reranking (Recommended: 10-20).

Top-N Reranked to Context5 chunks

Highest-scoring chunks supplied to the LLM context prompt (Recommended: 3-5).

Cosine Similarity Cutoff Threshold45%

Rejects chunks with similarity scores below this cutoff to prevent irrelevant context.

Hybrid Retrieval Balance70% Dense / 30% BM25

Ratio between semantic dense vector search and keyword/exact alarm code search.

Control LLM model selection and generation determinism for technical diagnostic instructions.

Default Reasoning ModelUsed for primary procedural synthesis and troubleshooting answers.

Temperature (Determinism)0.10

Lower values (0.05-0.15) enforce strict technical accuracy without hallucinations.

MEND-X natively processes queries in 10 languages for global field operators.

Preferred Output Language

Auto-detect Input Query LanguageAutomatically replies in the exact language used by the technician in the prompt.

Protect operators from dangerous or unverified equipment repairs.

Enforce Strict Industrial GuardrailsBlocks prompt override patterns, system prompt extraction, and out-of-domain conversational attempts.

Require Grounding & CitationsMandates that every actionable repair step links to an indexed manual section and page number before display.