LLM Integration UK
— Configure Your API
for AI Use

Most businesses asking about AI integration assume they need a complete rebuild. In reality, if you already have a working API, adding LLM capabilities is often a matter of adding three things: a function-calling schema, an embedding pipeline, and a prompt management layer.

At Niobotics, based in Leigh, Manchester, we specialise in retrofitting existing systems for AI use — without breaking what already works.

Step 1 — Define your function-calling schema

GPT-5 and Claude support tool use (formerly function calling) — the model can call your API endpoints as if they were functions. To enable this, you provide the LLM with a JSON schema describing what each endpoint does, what parameters it accepts, and what it returns.

For example, a product search endpoint becomes:

We build these schemas to match your existing OpenAPI spec, so the same documentation that describes your API to developers also describes it to AI models.

Step 2 — Build a RAG pipeline

Retrieval-Augmented Generation (RAG) lets your AI answer questions based on your own data — not just training data. The architecture is:

• Your documents/data are chunked and converted to vector embeddings using the OpenAI Embeddings API
• Embeddings are stored in pgvector — a PostgreSQL extension for vector similarity search
• When a user asks a question, the query is embedded and the most semantically similar chunks are retrieved
• Those chunks are injected into the LLM's context window as grounding data
• The LLM answers based on your data, not hallucinated training data

We implement this as a set of versioned API endpoints that sit alongside your existing routes.

Step 3 — Prompt injection defence

When users can influence what goes into an LLM prompt, they can attempt prompt injection — trying to override your system instructions. We implement:

• Input sanitisation to strip known injection patterns
• Separate system and user message contexts that can't be overridden
• Output validation to check the LLM's response matches expected schemas before returning it to users
• Rate limiting on AI endpoints to prevent abuse

Step 4 — Context window management

LLMs have finite context windows. For applications with long conversation histories or large document sets, naive approaches quickly hit token limits. We implement:

• Message summarisation to compress old conversation turns
• Semantic retrieval to only inject the most relevant RAG chunks
• Token counting middleware that trims context before it exceeds the model's limit

Cost

Adding LLM integration to an existing API at Niobotics costs around £100 — including the function-calling schema, RAG pipeline, prompt injection defences, and streaming response handler. Free maintenance is included.