RAG Chatbot Developer
Most chatbots in 2026 either hallucinate or refuse to answer. A properly built RAG (Retrieval-Augmented Generation) chatbot does neither — it pulls the relevant passages from your own documents and grounds the model in them, with clickable citations the user can verify. I am Ignacio (IGNAX), a Spain-born, Paraguay-based, solo full-stack developer + SEO/AEO specialist, and RAG chatbots are one of my most frequent builds for SaaS support, internal knowledge bases, and lead-qualification flows.
Who is this service for?
- SaaS companies whose support team answers the same 50 questions every week.
- Internal-knowledge-base owners whose Notion or Confluence has become unsearchable.
- Founders who want lead qualification done by a bot that actually knows the product.
- Agencies that need a senior RAG specialist on a white-label engagement.
If you are still trying to decide between RAG and fine-tuning, read RAG vs fine-tuning chatbots first — most teams do not need a fine-tuned model.
What is included in a RAG chatbot build?
- Discovery + one-page spec including corpus, target query volume, and citation requirements.
- Ingest worker that walks your source (Notion, Drive, Confluence, Zendesk, GitHub markdown, custom DB) and chunks documents intelligently (semantic-aware, not naive 500-token windows).
- Vector store on pgvector (Postgres extension) — no second database.
- Retrieval layer with hybrid search (BM25 + vector) and a reranker for hard queries.
- Generation layer on OpenAI or Anthropic with grounding-strict prompts.
- Citation rendering — every answer links back to the source passage.
- Feedback loop — thumbs up/down captured to a feedback table, used to grow the eval set.
- Eval harness — 20–50 canonical questions rerun on every prompt change, regression score reported.
- Frontend widget — embeddable widget (Svelte + Web Components) or a full chat UI, your call.
- Production deploy on your own cloud account. You own the code, the data, the embeddings.
- 14-day post-launch fix window included.
How does the build process work?
- Discovery call (free, 30 minutes). Bring a sample of the corpus and 10 questions you want the bot to answer well. I diagnose corpus health before quoting.
- One-page spec + fixed quote inside 48 hours. Corpus cleaning is quoted separately if needed.
- Ingest + retrieval prototype (week 1). Before any UI, I prove the retrieval layer answers your 10 canonical questions correctly.
- Generation + UI (week 2–3). Frontend widget, citation rendering, feedback capture.
- Eval pass + handover. 60-minute handover call, written runbook, eval harness committed.
Working hours: Mon–Fri, 8:00–17:00 PYT (GMT-3). Weekends off. Overlap with US East 7:00–16:00 EST, with Madrid 13:00–22:00 CET.
What stack do you ship with and why?
- Python 3.12 + LangChain (or LlamaIndex) — battle-tested abstractions for ingest, chunking, retrieval, and the reranker layer.
- pgvector on Postgres (Supabase or Neon) — one database, native HNSW indexes, free at MVP scale. See the pgvector documentation for the canonical reference.
- OpenAI or Anthropic — default to managed models. Anthropic for grounding-strict use cases; OpenAI for high-volume support.
- Cohere or BGE reranker — boosts retrieval precision on long-tail queries by ~10–15 points without retraining.
- SvelteKit or React widget — embeddable, accessibility-tested, supports streaming responses.
- Inngest or RQ — for scheduled corpus refresh and async eval runs.
If you are committed to a JavaScript-only stack, Vercel AI SDK + LangChain.js is a clean alternative — I can match.
How much does a RAG chatbot cost?
Pricing band: $2,500–$8,000 USD (Gs 15.3M–49.1M). The lower half covers a single-corpus chatbot on a clean source (e.g., your existing FAQ + product docs) with an embeddable widget. The upper half covers multi-source ingest (Notion + Drive + Zendesk), corpus cleaning, custom retrieval tuning, or a full standalone chat UI. Hourly outside that band: $30–$70 USD/hr. PYG values approximate, recalculated quarterly. Rate as of 2026-05-27: Gs 6,136.82 / USD. For the tradeoff analysis see RAG vs fine-tuning chatbots.
For agencies — white-label RAG chatbots
White-label RAG is a common subcontract because most agencies do not yet have an in-house ML specialist, and clients are asking. Day-rate option: $240–$560 USD/day under your brand and NDA. I sign your paperwork, deliver under your domain, and stay invisible to the end client. EN/ES bilingual coverage included.
What does the launch look like?
I deploy into your own cloud account — AWS, Cloudflare, Fly.io, or Render. You own the GitHub repo from commit one. After launch I run a 60-minute handover covering deploys, secrets, the embeddings refresh job, the eval harness, and how to read the citation feedback. A written runbook lives in the repo. Two weeks of free in-scope fixes follow.
What about ongoing maintenance?
Most RAG chatbots stabilize after the first month. When tuning is needed (corpus drift, new doc types, new eval questions), I offer a 5–10 hour/month bucket at 15% off hourly. No lock-in, 30-day rolling cancel. For broader background see what is AEO? — the same grounding principle (cite your sources, answer in the first 200 pixels) applies to both search and chat.
Related case studies and articles
- AI Support Chatbots — multi-client RAG bots grounded in client knowledge bases.
- AI Lead Generation Platform — RAG-powered lead qualification inside a SaaS.
- RAG vs fine-tuning chatbots — which one your project actually needs.
- What is AEO? — why answer-engine principles apply to chatbot design.
Ready to talk? Email hello@ignax.dev with a description of your corpus and the top 5 questions you want answered. I reply within one business day.
Working hours: Mon–Fri, 8:00–17:00 PYT (GMT-3). Weekends off. Overlap with US East 7:00–16:00 EST, with Madrid 13:00–22:00 CET.
Pricing
PYG values approximate, recalculated quarterly. Rate as of 2026-05-27: Gs 6,136.82 / USD.
For agencies
Available as a white-label senior subcontractor. Standard NDAs welcome. Day-rate option: $240–$560 USD/day depending on stack.
Email hello@ignax.devFrequently asked questions
What's the difference between a RAG chatbot and ChatGPT with file uploads?
ChatGPT with file uploads dumps your document into the context window of one conversation. A RAG chatbot indexes thousands of documents into a vector database, retrieves the relevant passages on every query, and grounds the model's answer in those passages — with citations. RAG scales to your full knowledge base, works for every user simultaneously, and exposes a measurable feedback loop. It is also the right pattern when you need answers to cite the source paragraph by name.
How accurate are the answers?
Honest answer: 'accurate enough to ship' is a function of corpus quality, chunking, retrieval, and reranking — not model choice. On a clean corpus (curated FAQs, structured docs) I see >90% acceptable answers in user evals. On a messy corpus (forum dumps, raw email threads) it drops to 60–70% without preprocessing. I quote retrieval-quality work as a separate line item in the spec when the corpus needs cleaning.
Does the chatbot hallucinate?
Less than a raw ChatGPT call, never zero. I mitigate with three controls. First, the model is instructed to answer only from retrieved passages and to say 'I do not know' when retrieval is empty. Second, every answer renders inline citations the user can click to verify. Third, a weekly eval set (20–50 canonical questions) is rerun on every prompt change to catch regressions before deploy.
Should I use OpenAI or Anthropic?
For RAG specifically, Anthropic Claude tends to follow grounding instructions more strictly — it is more willing to say 'I do not know' when retrieval is weak. OpenAI's GPT-4.1 is faster and cheaper for high-volume support flows. I default to Claude for legal, medical, or compliance-sensitive corpora; OpenAI for B2C support chatbots where speed beats caution. Model is an env var either way.
Where does the vector database live?
On the same Postgres database as the rest of your app, via the pgvector extension. There is no second database to operate. Supabase and Neon both ship pgvector by default. I only reach for Pinecone, Weaviate, or Qdrant when you have over ten million vectors or sub-50ms latency requirements — almost never for SMB use cases.
How do I keep the chatbot's knowledge base in sync?
Through a scheduled ingest job. Every night (or on every doc update, via webhook) the worker walks your source — Notion, Google Drive, Confluence, Zendesk, GitHub markdown, whatever — diffs against the vector store, and updates only the changed chunks. Stale chunks are flagged in the cost dashboard so you can see when your corpus drifts. The full ingest playbook is documented in the runbook.
Can you white-label this for our agency?
Yes. White-label RAG chatbots are a common subcontract — clients ask for them, agencies often do not have the in-house specialist. Day-rate $240–$560 USD/day, NDA-friendly, deliverables under your brand. I stay invisible to the end client. EN/ES bilingual coverage included at no extra cost.
What happens after launch?
Two weeks of free in-scope bug fixes are included. After that, RAG chatbots typically need a 5–10 hour/month bucket for corpus refresh tuning, eval-set updates, and the occasional prompt drift. I do not push retainers — most chatbots stabilize after month one. When you do need work I bill against a 15% discounted maintenance rate, 30-day rolling cancel.