Build a coding agent (RAG)

Give a chat model knowledge of your codebase: embed the repo, retrieve the most relevant chunks for a question, and feed them to the model. Embeddings and chat run on the same AIx key and bill to one wallet.

1 · Set up the client

from openai import OpenAI

client = OpenAI(base_url="https://api.aix.theaimart.co/v1", api_key="$AIX_KEY")
EMBED = "BAAI/bge-large-en-v1.5"
CHAT  = "anthropic/claude-sonnet-4-6"   # or deepseek-ai/DeepSeek-V3.2, Qwen/Qwen3-Coder-...

2 · Chunk & embed the repo

import os, glob

def chunks(text, size=1200, overlap=150):
    for i in range(0, len(text), size - overlap):
        yield text[i:i + size]

docs = []   # [(path, chunk_text)]
for path in glob.glob("src/**/*.*", recursive=True):
    try:
        text = open(path, encoding="utf-8").read()
    except Exception:
        continue
    for c in chunks(text):
        docs.append((path, c))

# Embed in batches (max 100 inputs per request)
vectors = []
for i in range(0, len(docs), 100):
    batch = [c for _, c in docs[i:i + 100]]
    res = client.embeddings.create(model=EMBED, input=batch)
    vectors.extend(d.embedding for d in res.data)

3 · Retrieve the top matches

import numpy as np

mat = np.array(vectors)
mat /= np.linalg.norm(mat, axis=1, keepdims=True)

def search(query, k=6):
    q = client.embeddings.create(model=EMBED, input=query).data[0].embedding
    q = np.array(q); q /= np.linalg.norm(q)
    scores = mat @ q
    return [docs[i] for i in scores.argsort()[-k:][::-1]]

4 · Answer with retrieved context

def ask(question):
    hits = search(question)
    context = "\n\n".join(f"# {path}\n{chunk}" for path, chunk in hits)
    r = client.chat.completions.create(
        model=CHAT,
        messages=[
            {"role": "system", "content": "Answer using ONLY the provided code context. Cite file paths."},
            {"role": "user", "content": f"Context:\n{context}\n\nQuestion: {question}"},
        ],
    )
    return r.choices[0].message.content

print(ask("Where do we verify the API key?"))

Cost & latency tips

• Embed once, cache the vectors — re-embed only changed files.
• Retrieval keeps the prompt small → cheaper + faster than dumping the whole repo.
• Use a cheaper chat model for routine Q&A and a stronger one for hard refactors — switch by changing the CHAT string.

Next steps

• Tool & function calling — let the agent run actions
• Streaming a chat UI — show answers token-by-token