{"id":43,"date":"2026-10-09T19:51:54","date_gmt":"2026-10-09T11:51:54","guid":{"rendered":"http:\/\/192.168.31.81\/index.php\/faq-en\/"},"modified":"2026-10-09T19:51:54","modified_gmt":"2026-10-09T11:51:54","slug":"faq-en","status":"publish","type":"page","link":"https:\/\/moltbot.yundulizhan.cn\/index.php\/faq-en\/","title":{"rendered":"FAQ on Enterprise AI Adoption"},"content":{"rendered":"

These are the questions companies ask most when putting AI into production. Each answer is meant to be concrete and verifiable, not generic.<\/p>\n

\nHow do we build an enterprise AI knowledge base?<\/summary>\n

Separate "knowledge" from "documents". Documents are the PDFs, Word files and policy compilations; knowledge is the smallest unit inside them that can be retrieved and traced back to its source. The core work is: de-duplicate, determine versions, cut to the right granularity, and define refusal boundaries. Dumping PDFs straight into a vector store is the same as doing nothing.<\/p>\n<\/details>\n

\nWhy does the model answer inaccurately?<\/summary>\n

In my experience 80% of cases are not a model problem but one of these three: retrieved fragments are incomplete or the wrong version, nobody defined what "correct" means, and nobody maintains it after launch. The remaining 20% is genuine model capability limits.<\/p>\n<\/details>\n

\nHow do we improve poor RAG results?<\/summary>\n

Do not rush to swap the model. Check in order: is the chunking granularity right, are the retrieved fragments right, does the prompt require "say you do not know when unsure", is there a version conflict. These four steps usually fix most problems; changing the model is the last resort.<\/p>\n<\/details>\n

\nWhat hardware does a private deployment need?<\/summary>\n

It depends on model size. 7B\u20138B at Q4 needs about 5GB of VRAM, so a 12GB card is enough; 14B at Q4 needs about 9GB \u2014 a 12GB card can run it but with limited context. For internal Q&A in a mid-sized company, 7B\u201314B is usually enough; you do not need tens of billions of parameters on day one.<\/p>\n<\/details>\n

\nHow large a model fits in 12GB of VRAM?<\/summary>\n

14B at Q4_K_M (about 9GB) runs, with the rest left for context \u2014 keep the context window under 8K. For longer context, drop to 7B.<\/p>\n<\/details>\n

\nWho is responsible when the AI agent answers wrong?<\/summary>\n

This has to be written into the contract. My approach: define the "must refuse" boundary, route anything beyond it to a human, and log every answer so it can be traced back. Assigning responsibility starts with knowing when and what it answered, so logging is standard.<\/p>\n<\/details>\n

\nIs AI Q&A compliant in a hospital?<\/summary>\n

Compliance starts with a clear role: it may only do information lookup and process guidance, never diagnosis, never replace a doctor\u2019s judgement. Anything about a specific condition must be refused and redirected to a doctor. I hard-code that boundary when building the knowledge base.<\/p>\n<\/details>\n

\nHow much does enterprise AI cost?<\/summary>\n

It depends on how messy your knowledge assets are. The health check is free, and after it I can quote accurately. Order of magnitude: building a base for a medium-complexity scenario usually runs from tens of thousands upward, with ongoing tuning billed monthly. Any quote given without a check is not trustworthy.<\/p>\n<\/details>\n

\nHow is AI performance evaluated?<\/summary>\n

Sample 50\u2013100 real business questions, have humans label the reference answers, then run four metrics: factual accuracy, citation hit rate, should-refuse rate, and manual review volume. The key is that these must be reproducible \u2014 someone else running it gets the same numbers.<\/p>\n<\/details>\n

\nOur material is a mess \u2014 should we tidy it up first?<\/summary>\n

No. Tidying is part of my job, and I know what structure makes it retrievable. If you tidy it yourself first, it will most likely need redoing.<\/p>\n<\/details>\n

\nLocal deployment or cloud API?<\/summary>\n

Look at data sensitivity. Internal policies, client files, unpublished expertise \u2014 local. General queries on public information \u2014 cloud API is cheaper. A mix is common: sensitive knowledge retrieved locally, general understanding handled in the cloud.<\/p>\n<\/details>\n

\nHow soon do we see results?<\/summary>\n

The health check report lands in 2\u20133 days and tells you where the problems are straight away. After 4\u20136 weeks of base building, the metrics improve clearly. But staying accurate is a long-term job that needs monthly maintenance.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

These are the questions companies ask most when putting […]<\/p>\n","protected":false},"author":0,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-43","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/pages\/43","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/types\/page"}],"replies":[{"embeddable":true,"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/comments?post=43"}],"version-history":[{"count":0,"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/pages\/43\/revisions"}],"wp:attachment":[{"href":"https:\/\/moltbot.yundulizhan.cn\/index.php\/wp-json\/wp\/v2\/media?parent=43"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}