In-House GPT Knowledge Assistant
Budget: €1,500 – €3,000 EUR
I want to set up a fully self-hosted GPT assistant that lives on our own Intel NUC or a similar on-premise server. The goal is to turn our existing company knowledge—PDF-Dokumente, Excel-Tabellen und einfache Textdateien—into a searchable, conversational resource that every employee can tap into, either by typing or by speaking.
Scope of work
• Model & infrastructure – Select or fine-tune an LLM that offers a strong price-performance ratio, install it locally, and configure GPU / CPU use so it runs reliably inside our network.
• Data ingestion – Build a pipeline that continuously reads the above file types, creates embeddings, and stores them in a local vector database (e.g. Chroma, FAISS, or similar).
• Chat & voice interface – Deliver a lightweight web UI plus speech-to-text and text-to-speech hooks so the assistant answers both text and voice queries. Access should work for all employees behind our firewall without external calls.
• Security & admin – Provide role-based controls, logging, and a straightforward way for us to add new documents or retrain the model.
• Knowledge transfer – Hand over clear documentation and a short remote walkthrough so we can maintain and expand the system ourselves.
Acceptance criteria
1. The assistant returns accurate answers sourced from our internal files within acceptable latency on the provided hardware.
2. Text and voice interactions function smoothly in a test with multiple employees.
3. No data leaves our network during operation; all traffic and storage stay internal.
If you have hands-on experience with local LLM deployment, RAG pipelines, Whisper or similar speech modules, and secure enterprise roll-outs, I would be glad to review your approach and previous work.
Scope of work
• Model & infrastructure – Select or fine-tune an LLM that offers a strong price-performance ratio, install it locally, and configure GPU / CPU use so it runs reliably inside our network.
• Data ingestion – Build a pipeline that continuously reads the above file types, creates embeddings, and stores them in a local vector database (e.g. Chroma, FAISS, or similar).
• Chat & voice interface – Deliver a lightweight web UI plus speech-to-text and text-to-speech hooks so the assistant answers both text and voice queries. Access should work for all employees behind our firewall without external calls.
• Security & admin – Provide role-based controls, logging, and a straightforward way for us to add new documents or retrain the model.
• Knowledge transfer – Hand over clear documentation and a short remote walkthrough so we can maintain and expand the system ourselves.
Acceptance criteria
1. The assistant returns accurate answers sourced from our internal files within acceptable latency on the provided hardware.
2. Text and voice interactions function smoothly in a test with multiple employees.
3. No data leaves our network during operation; all traffic and storage stay internal.
If you have hands-on experience with local LLM deployment, RAG pipelines, Whisper or similar speech modules, and secure enterprise roll-outs, I would be glad to review your approach and previous work.