We manage hundreds of PostgreSQL, MSSQL, Oracle, MongoDB, and MySQL servers across multiple products. Every day, engineers ask the same kinds of questions:<\/p>\n\n\n\n
> “How many critical alerts came in last week?”<\/p>\n\n\n\n
> “Show me the database growth trend for server X.”<\/p>\n\n\n\n
We needed a chatbot that could answer these questions and get smarter<\/em> over time based on our team’s feedback. But we didn’t have access to GPU clusters for fine-tuning, and we didn’t want to send proprietary data to external APIs.<\/p>\n\n\n\n The constraint: Build an AI assistant that learns from users, runs entirely on-prem(with very basic server config is more than enough), and costs nothing to train.<\/p>\n\n\n\n The Architecture<\/strong><\/p>\n\n\n\n Here is the complete system at a glance:<\/p>\n\n\n\n How the Chatbot Works -Three Layers<\/p>\n\n\n\n The chatbot does not just throw everything at an LLM. It uses a layered approach to answer questions as fast and accurately as possible:<\/p>\n\n\n\n Layer 1: Rule Engine (< 50ms): Most operational questions follow patterns. “Show critical alerts for PROJ1<\/strong>” matches a rule, fills a SQL template, and returns data instantly. No LLM needed.<\/p>\n\n\n\n Layer 2: Knowledge Base: If the rule engine doesn’t match, we check reviewer notes and historical incident data for relevant context.<\/p>\n\n\n\n Layer 3: LLM with RAG: For everything else, the local LLM takes over. But before it generates an answer, we retrieve similar past Q&A pairs rated positively by users and inject them into the prompt.<\/p>\n\n\n\n The Learning Loop: How RAG Makes the Chatbot Smarter<\/strong><\/p>\n\n\n\n Every response the chatbot gives has thumbs-up\/down buttons. When a user interacts with these buttons, a learning cycle begins:<\/p>\n\n\n\n What happens on \ud83d\udc4d (Thumbs Up)<\/p>\n\n\n\n The question-answer pair is promoted to the Learned Q&A table, a curated knowledge base of verified good answers. If the same question gets multiple thumbs-up, the number quality_score increases, making it rank higher in future retrievals.<\/p>\n\n\n\n What happens on \ud83d\udc4e + Correction<\/p>\n\n\n\n The user types what the answer should<\/em> have been. The correction replaces the bad answer in the knowledge base. Next time a similar question is asked, the LLM sees the corrected answer as context.<\/p>\n\n\n\n The key insight<\/strong><\/p>\n\n\n\n The LLM model weights never change. We don’t fine-tune. Instead, we give the LLM a “cheat sheet” of verified answers before it responds. This is Retrieval-Augmented Generation, the same architecture behind many production AI systems, adapted for a small local model.<\/p>\n\n\n\n Pure-SQL RAG:- No Vector Database Required<\/p>\n\n\n\n Most RAG implementations require a vector database (Pinecone, pgvector, Weaviate) and an embedding model. We didn’t have access to either.<\/p>\n\n\n\n Instead, we built RAG using PostgreSQL’s built-in pg_trgm extension for fuzzy text matching:<\/p>\n\n\n\n How it works:<\/strong><\/p>\n\n\n\n 1. pg_trgm breaks text into 3-character sequences (trigrams)<\/p>\n\n\n\n 2. It compares the user’s question against every learned Q&A<\/p>\n\n\n\n 3. Matches above 15% similarity and minimum quality score are returned<\/p>\n\n\n\n 4. Top 3 matches are formatted and appended to the LLM’s system prompt<\/p>\n\n\n\n No embeddings. No vector DB. No GPU. Just PostgreSQL doing what it does best.<\/p>\n\n\n\n What the LLM Actually Sees<\/strong><\/p>\n\n\n\n When a user asks “critical alerts this week”<\/em>, the LLM receives this prompt:<\/p>\n\n\n\n The LLM reads the verified answer as context and produces a response that’s informed by what worked before. Over time, as more feedback accumulates, the context gets richer and more accurate.<\/p>\n\n\n\n Key Metrics<\/strong><\/p>\n\n\n\n We track four metrics to verify the system is actually learning:<\/p>\n\n\n\n learned_qa_count<\/strong> – New verified answers added to knowledge base<\/p>\n\n\n\n use_count<\/strong> – Learned answers being retrieved and used in prompts<\/p>\n\n\n\n quality_score<\/strong> – Same answer reinforced by multiple users<\/p>\n\n\n\n corrections<\/strong> – Users actively teaching better answers<\/p>\n\n\n\n After just a running few learning cycles, we saw:<\/p>\n\n\n\n – 3 learned Q&A pairs with quality scores up to 1.5 (multi-user reinforcement)<\/p>\n\n\n\n – Learned answers being retrieved 3+ times for similar queries<\/p>\n\n\n\n – 52-71% similarity matches catching question variations accurately<\/p>\n\n\n\n What’s Next<\/p>\n\n\n\n 1. More feedback = better retrieval.<\/p>\n\n\n\n 2. Full-text search with ts_vector.<\/p>\n\n\n\n 3. Fine-tuning.<\/p>\n\n\n\n Key Takeaway<\/strong><\/p>\n\n\n\n With PostgreSQL’s pg_trgm, a local 7B model, and a simple feedback \u2192 store \u2192 retrieve \u2192 inject loop, you can build a domain-specific assistant that gets smarter with every interaction.<\/p>\n\n\n\n The learning happens not by changing the model, but by changing what the model sees.<\/p>\n\n\n\n Note: This is something I tried over the last few days. Since access to GPUs and very large models isn\u2019t always feasible (cost + infra constraints), I wanted to see how far we can go with a smaller local model + smart retrieval. Curious to hear what others think, feedback is very welcome \ud83d\ude42.<\/p>\n\n\n\n <\/p>\n","protected":false},"excerpt":{"rendered":" How I built a domain-specific chatbot that learns from every user interaction using RAG, PostgreSQL, and a local LLM, with zero training cost :-). The Problem We manage hundreds of PostgreSQL, MSSQL, Oracle, MongoDB, and MySQL servers across multiple products. Every day, engineers ask the same kinds of questions: > “How many critical alerts came […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10,32],"tags":[35,33,34],"class_list":["post-292","post","type-post","status-publish","format-standard","hentry","category-ai","category-chatbots","tag-database-operations","tag-rag","tag-self-leaning-ai-chatbot"],"_links":{"self":[{"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/posts\/292","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/comments?post=292"}],"version-history":[{"count":1,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/posts\/292\/revisions"}],"predecessor-version":[{"id":293,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/posts\/292\/revisions\/293"}],"wp:attachment":[{"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/media?parent=292"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/categories?post=292"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/metaintelligencesystems.com\/blog\/wp-json\/wp\/v2\/tags?post=292"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\n