« We plug an AI into your PDFs » hides a dozen very different techniques, with known strengths and pitfalls. Here is the landscape as the research describes it, then the principles that guide our choices — not the exact recipe, but the why.
| Technique | What it does | Its limit |
|---|---|---|
| Keyword search (BM25) | Retrieves passages that share the exact words. Fast, exact on codes (« RR42 », « 45 Nm »). | Blind to synonyms and to field paraphrases. |
| Vector search (embeddings) | Retrieves by meaning: « it won't build up pressure » finds « pressure loss » even without a common word. | May miss an exact code match; depends on the model. |
| Hybrid search (fusion) | Combines the two and merges the rankings. The research widely recommends it as a robust foundation. | Requires careful tuning of the fusion. |
| Reranking | Reorders the candidates according to their true relevance to the question. | Compute cost; useless if the initial recall is bad. |
| Citation verification (grounding) | Checks that every claim is supported by a source before displaying it. This is what makes hallucinations drop. | One extra verification call; requires good sources. |
| GraphRAG (knowledge graph) | Builds a graph of entities to reason over complex links. | Heavy to build and maintain; often oversized. |
| Fine-tuning the model | Retrains the model on your data. | Costly, freezes the knowledge, does not cite its sources, goes stale. |
Our exact sequence — the order, the thresholds, the settings — is our engineering work, and we don't publish it. But the principles that govern it are transparent. They all stem from the same obsession: a wrong answer costs more than no answer.
A technician writes « it's getting hot » but your manual says « bearing overheating » — and they also type « RR42 » which must match to the character. Neither the keyword alone nor the vector alone is enough. We combine, as the research recommends.
This is the principle that most distinguishes us from a « chatbot on PDF »: an answer not supported by a source is not displayed. It's exactly the mechanism that the literature associates with the highest rates of citation accuracy.
When a case falls outside what's documented, the assistant tells you so and turns that moment into a card to be validated. Research on system reliability confirms it: answering where you are sure makes a system more trustworthy.
For documentation-based troubleshooting, we favor a robust hybrid foundation and strict verification — effective from day one, simple to operate. Knowledge-graph approaches remain in reserve for the cases that will truly justify them, when the complexity of the links demands it.
Our approach reads your documents live: every answer cites the page and stays current with every update of your manuals, and your data trains no model. That's what keeps the exact citation and the isolation between clients — where retraining a model would freeze it and make it lose the link to the source (documented trade-off).