Multi-Tenant RAG Dashboard for Rapid Client Onboarding with Custom Knowledge Bases
Engineered a production-deployed MVP that lets businesses create OpenAI Assistants, ingest documents from uploads or Google Drive, and chat with a knowledge-aware agent from a single dashboard—built for repeatable rollout across multiple client datasets.
"From concept to deployed RAG SaaS MVP in weeks—complete with Drive ingestion, assistant management, and spam-safe access control"
Frontend
Backend
AI Models
Infrastructure
< 1 month
MVP delivery timeline
Aligned to client target; delivered deployed dashboard + RAG ingestion within the agreed maximum timeframe.
16 (down from 20)
Reduced retrieval payload
Limited retrieved chunks to prevent \"maximum TPM exceeded\" errors and stabilize responses under token/rate limits.
$90/year saved
Avoided paid SSL certificate
Implemented Let's Encrypt with auto-renew on VPS instead of purchasing a commercial wildcard/subdomain SSL plan.
5–10 USD
Low-cost development runtime
Client-provided OpenAI balance used for development/testing as planned.
The Challenge
Problem Statement
The client wanted to commercialize a repeatable RAG offering: one codebase that could onboard multiple companies, securely ingest their internal files (PDF/DOCX/TXT/JSON and Google Drive content), and provide a professional chat experience. Key risks included data privacy expectations, ingestion at scale (large Drive folders), OpenAI vector store quirks in beta, and controlling usage costs/spam before payments and full auth were ready.
The Solution
Our Approach
Architected a dashboard-first MVP that wraps the OpenAI Assistants workflow into a self-serve admin portal. The system enables creating assistants with configurable model/tooling (including Code Interpreter), uploading and indexing documents into an OpenAI Vector Store, and chatting through a web UI. Added Google Drive ingestion (OAuth-based approach planned via Google Cloud app ownership) and deployed the app to a VPS behind a custom subdomain with HTTPS to satisfy Google OAuth and secure transport requirements.
AI Implementation
Document-Aware Assistants (RAG) with Vector Store + File Search
Technical Details
Implemented an OpenAI Assistants-based RAG pipeline using GPT-4o with File Search over an OpenAI-hosted vector store. Tuned retrieval behavior to reduce token pressure and avoid rate-limit failures (reduced retrieved results from 20 to 16 and added error tolerance around File Search). Added support for multiple file types including JSON (telegram exports) and handled edge cases where response-format settings (JSON mode vs AUTO) impacted assistant replies.
Business Value
Enabled fast, repeatable onboarding for multiple companies—upload a new dataset or connect Drive, generate an assistant, and immediately provide accurate, context-grounded answers. This reduced manual setup work and positioned the product as a scalable \"RAG-as-a-service\" offering for beta clients.
Technical Deep Dives
Challenges We Solved
Google Drive integration blocked by missing HTTPS
OAuth-based Google Drive connectivity failed when the dashboard was served over plain HTTP, preventing secure redirects and API permission flows.
Deployed HTTPS on the VPS and secured the subdomain to meet OAuth/security requirements. Implemented certificate automation via Let's Encrypt to eliminate annual renewals and keep the environment production-ready.
Let's EncryptNginxNamecheap VPSDNS/Subdomain Routing
OpenAI Vector Store beta ingestion failure on specific JSON export
A particular file (\"Results (1).json\") caused ingestion/search instability; behavior reproduced even inside OpenAI Playground, indicating a platform-side issue.
Triaged by isolating the problematic file, validating behavior in OpenAI Playground, and reporting the issue to OpenAI. Implemented workaround strategies: improving File Search error tolerance, advising alternative formats/exports, and confirming other JSON files indexed correctly.
OpenAI Assistants PlaygroundOpenAI Vector StorePythonVS Code
Assistant response failure due to misconfigured response format
Assistants appeared \"broken\" after enabling JSON response format, causing unexpected behavior during chat and retrieval testing.
Identified the root cause (advanced setting set to JSON instead of AUTO) and provided a corrective workflow: recreate/update assistant with response format set to AUTO; reserved JSON mode for future API-oriented assistant outputs.
OpenAI Assistants APIDashboard Configuration UI
Post-deploy regression in Drive/upload flow
After pushing an updated build, Google Drive integration and uploads stopped working due to a deployment mistake.
Hotfixed the deployment quickly and restored functionality, demonstrating production support readiness even after contract close.
PythonVPS DeploymentStreamlit App Runtime
Process
Project Timeline
1
Discovery
Clarified SaaS-style requirements: multi-company reuse, document ingestion (PDF/DOCX/TXT/JSON), OpenAI Assistants with embeddings + vector database, and a professional dashboard experience. Aligned constraints around privacy expectations, security (TLS), and beta testing without payments.
2
Build
Built a full web dashboard to create/manage assistants, upload and index documents, enable tool selection (e.g., Code Interpreter), and provide a chat interface. Implemented Google Drive ingestion workflow planning (Google Cloud app ownership + user consent) and iterated based on client testing, including fixes for response-format misconfiguration and vector store edge cases.
3
Launch
Provisioned Namecheap VPS, configured DNS/subdomain (dashboard.celestix.ai), deployed the app, and secured it with HTTPS using Let's Encrypt. Added basic access gating (password) and delivered source code + runbook (README) for handover; continued short post-launch bugfix support.
Ready to Build Something Similar?
Let's discuss how we can help transform your business with AI.
Start Your Project
