RAG Chatbots · Osaka
RAG Chatbot Development — AI Trained on Your Internal Documents
"Our manuals are so vast that new hires can't find anything," "It takes half a day to find a similar past case in our meeting minutes," "It takes 30 minutes to pull the right page out of a sales deck" — a RAG (Retrieval-Augmented Generation) chatbot is a perfect fit for these "let AI search our internal documents" needs.
Sugureru builds AI chatbots that let an LLM search your internal manuals, FAQs, meeting minutes, contracts, and past-project data, then answer employees' natural-language questions with sources cited. Rather than the multi-million-yen scale of major system integrators, our hallmark is lightweight implementation for SMEs that you can start with a PoC from ¥500,000.
Do any of these challenges sound familiar?
- ▸"Not all employees can find the right manual or FAQ, so they end up asking a veteran out loud anyway."
- ▸"Knowledge in our meeting minutes and past projects is locked in individuals — when a veteran leaves, it disappears with them."
- ▸"We spend 1–2 hours every day searching through sales decks and proposal templates."
- ▸"ChatGPT on its own doesn't know our internal information, so it isn't usable for actual work."
- ▸"For security reasons, we want to avoid sending confidential documents to an external LLM API."
Three reasons clients choose Sugureru
1. Lightweight implementation for SMEs from ¥500,000
Market quotes for RAG builds typically start at ¥5,000,000 even for small projects, and run ¥10,000,000–30,000,000 for mid-sized ones. Rather than the full custom builds of major system integrators, Sugureru delivers working implementations starting from a ¥500,000 PoC. We keep costs down by making full use of open-source technologies such as LangChain, LlamaIndex, Pinecone, Weaviate, and ChromaDB.
2. Private-network ready — a configuration that keeps data in-house is possible
Under the enterprise agreements for Azure OpenAI and Amazon Bedrock, your input data is not used for model training. And when security requirements are strict, a fully closed/private-network configuration that never sends any data to the cloud is also possible, using on-premise open-source LLMs such as Llama 3 or Mistral.
3. Any format — PDF, Word, Excel, even meeting audio
We support PDF (including scanned images), Word, Excel, PowerPoint, Markdown, HTML, and even meeting audio recordings (via automatic transcription). We can also set up automatic integration with your existing document-management tools such as SharePoint, Google Drive, Notion, and Confluence.
What a RAG chatbot can do
AI search of internal manuals
Train it on work rules, operational manuals, and equipment-operation procedures so it instantly answers employees' natural-language questions. It answers questions like "How many days of paid leave do I get?" with sources cited.
AI search of past projects and meeting minutes
Train it on past sales-meeting minutes, proposals, and contracts so an investigation like "What similar terms did we handle on Company A's project three years ago?" finishes in 30 seconds. It prevents the loss of veterans' tacit knowledge.
Automated FAQ responses for customer support
Train it on your past inquiry history and FAQs to automatically answer 70% of inquiries over LINE, email, or web chat. A hybrid operation that escalates only the difficult cases to an operator.
Search of sales decks and proposal templates
Meets the instant-answer needs of the sales floor, such as "Pull up the materials we used for the manufacturing inventory-management system proposal" or "What's the latest version of the price list?" It can be called from Slack or Teams.
Pricing plans
| Plan | Price | Timeline | What's included |
|---|---|---|---|
| PoC Lite | From ¥500,000 | 3–4 weeks | A working prototype for 1 domain with 100 documents or fewer |
| Full Deployment | ¥1,000,000–3,000,000 | 8–12 weeks | Multiple domains, private-network support, integration with existing tools |
| Full Operations | From ¥3,000,000 | Ongoing contract | Continuous training, model improvement, adding new domains |
All listed prices exclude tax. Monthly operating cost: ¥10,000–50,000/month (including LLM API usage fees and vector database hosting, for a scale of 500–2,000 documents).
How we work
-
01
Business interview and document review
We map out which tasks involve the "effort of searching for documents," and the types, counts, and storage locations of the target documents. Security requirements are confirmed here as well.
-
02
Architecture design and quotation
We propose a configuration and cost tailored to your requirements: cloud or private network, LLM selection (GPT / Claude / Llama), and vector database selection (Pinecone / ChromaDB / Qdrant).
-
03
PoC development and accuracy validation
We run a prototype on your actual internal documents and validate answer accuracy together with your reviewers. You can judge "whether it's actually usable" before committing to full deployment.
-
04
Production rollout and operation
We integrate it into whichever entry point is easiest for your team — Slack, Teams, LINE, a web UI, and more. Automatic ingestion of document updates and continuous improvement of answer accuracy are handled in the operation phase.
Use cases by industry
Representative use cases we are actually consulted on across Osaka and the Kansai region.
- Manufacturing: searching past drawings and specifications, searching quality-standard documents, turning production-equipment manuals into AI
- Professional services (lawyers, tax accountants, administrative scriveners): searching past cases and precedents, generating contract drafts, sharing the latest tax-law information internally
- Real estate: automated responses for property information + legal questions, searching past transaction examples
- IT / SaaS support: automatic routing of inquiries, automated Tier 1 support responses
- Education and training: searching textbooks and manuals, automated FAQ responses for new-hire training
Frequently asked questions
Which document formats do you support?
We support all major formats, including PDF, Word, Excel, PowerPoint, Markdown, HTML, plain text, and CSV. Scanned-image PDFs can be transcribed via OCR and used for training. We also support the flow of meeting audio recording → automatic transcription → training.
I'm worried that our internal information will be used to train an external LLM.
Under the enterprise agreements for Azure OpenAI and Amazon Bedrock, your input data is not used for model training (this is explicitly guaranteed in the contract). For especially confidential needs, on-premise open-source LLMs (such as Llama 3 or Mistral) are also an option. We can implement a configuration that never sends any data to the cloud at all.
How accurate are the AI's answers? How do you address hallucinations (false answers)?
By the very design of RAG, the AI can only answer from the internal documents it has retrieved in advance, so hallucinations are greatly reduced. We also implement source-cited answers that make clear which document and which page a quote came from, giving users a UI that lets them judge whether an answer is true.
Will it automatically retrain when documents are added or updated?
Yes. By integrating with document-management tools such as SharePoint, Google Drive, Notion, and Confluence, we can build a system that automatically detects document updates and retrains. A manual upload method is also available.
Can it be used from our existing chat tools such as Slack, Teams, or LINE?
Yes. We can integrate it with the tools your employees use every day, including a Slack Bot, Microsoft Teams Bot, LINE Bot, and an internal web chat UI. There's no need to make people learn a new tool, which leads to higher adoption.
Can you explain the relationship between the number of documents and the cost?
Initial build cost depends less on the number of documents than on the complexity of the target domain. Whether it's 100 documents or 1,000, the base configuration doesn't change as long as the domain is the same. Monthly operating cost consists of the vector database and LLM API usage fees, with a guideline of ¥10,000–50,000/month for a scale of 500–2,000 documents.
Considering a RAG chatbot? Start with a free consultation.
We welcome inquiries even at the stage where you're not yet sure "which of our documents would deliver results if trained into AI." First consultation and quotation are free, and we usually respond within 24 hours.