Custom AI Assistant Integration for IDE

Your Django team of 12 developers spends hours on repetitive code reviews. We built a custom AI assistant that understands your codebase and delivers relevant suggestions with <200ms latency. Here's how.

A custom AI assistant for the IDE is not just autocomplete on steroids. It keeps the entire project context: open files, change history, database schema, tests. A properly built assistant knows you're writing a user registration function in a Django project with PostgreSQL and suggests code compatible with your models and conventions.

Problems we solve

Generic models don't know project context. GitHub Copilot gives average-quality suggestions, ignoring internal APIs, custom ORM methods, and architectural decisions. The acceptance rate of such suggestions rarely exceeds 23%.

Code confidentiality. Teams with NDAs cannot send code to cloud services. A fully local stack is required.

Suggestion latency. Cloud solutions often have latency >500ms, killing the magic. For inline completion, latency <200ms is critical.

A custom assistant solves all three: it uses your codebase, works locally, and delivers suggestions in 80–150ms.

Architecture of an IDE assistant

A full Copilot-like assistant consists of several layers:

• Context Collector — gathers relevant context: current file, imports, related files, cursor position, selected code, clipboard.
• LSP Bridge — interacts with the Language Server Protocol to get AST, types, definitions.
• Retrieval Engine — semantic search over the codebase using embeddings (CodeBERT, text-embedding-3-small) and a vector store with RAG.
• LLM Gateway — request routing: fast model for inline completion, powerful model for chat/refactoring.
• Response Renderer — output formatting: diff for refactoring, ghost text for completion, markdown for chat.

Why a custom AI assistant outperforms GitHub Copilot

A custom assistant uses your project's context: codebase indexes, DB schemas, issue trackers. This yields more relevant suggestions than generic models. In our case study, acceptance rate rose from 23% to 41%, and subscription costs were cut in half (saving $2,000 per month). Plus, you have full data control — no code leaks to cloud services.

Continue.dev — open-source foundation

Continue.dev (https://github.com/continuedev/continue) is the most mature open-source alternative to GitHub Copilot. It supports VS Code and JetBrains, configurable via ~/.continue/config.json.

{
  "models": [
    {
      "title": "Claude 3.5 Sonnet",
      "provider": "anthropic",
      "model": "claude-sonnet-4-5",
      "apiKey": "$ANTHROPIC_API_KEY"
    },
    {
      "title": "Ollama Qwen2.5-Coder",
      "provider": "ollama",
      "model": "qwen2.5-coder:7b",
      "apiBase": "http://localhost:11434"
    }
  ],
  "tabAutocompleteModel": {
    "title": "Autocomplete",
    "provider": "ollama",
    "model": "qwen2.5-coder:1.5b"
  },
  "contextProviders": [
    {"name": "code", "params": {}},
    {"name": "docs", "params": {}},
    {"name": "diff", "params": {}},
    {"name": "terminal", "params": {}},
    {"name": "problems", "params": {}},
    {"name": "folder", "params": {}},
    {"name": "codebase", "params": {}}
  ],
  "slashCommands": [
    {"name": "edit", "description": "Edit highlighted code"},
    {"name": "comment", "description": "Write comments for the code"},
    {"name": "tests", "description": "Write unit tests"},
    {"name": "share", "description": "Export the chat session"}
  ]
}

Key feature: tabAutocompleteModel uses a fast local model (1.5B parameters), while chat uses a powerful cloud model. Inline completion latency: 80–150ms on Qwen2.5-Coder 1.5B via Ollama.

import { ContinueConfig, IContextProvider } from "@continuedev/core";

class DatabaseSchemaProvider implements IContextProvider {
  get description() {
    return { title: "db", displayTitle: "Database Schema", description: "Current database schema", type: "normal" };
  }

  async getContextItems(query: string, extras: any) {
    const schema = await fetchDatabaseSchema();
    return [{ name: "Database Schema", description: "Current DB schema", content: schema }];
  }
}

export function modifyConfig(config: ContinueConfig): ContinueConfig {
  config.contextProviders = [...(config.contextProviders || []), new DatabaseSchemaProvider()];
  return config;
}

How we configure context-aware suggestions for your project

The setup process consists of four steps.

1. Codebase analysis: we identify key patterns, internal APIs, and database structure. We use a static analyzer to extract metadata.

2. Custom context providers: for each source (DB schema, Jira, documentation) we write a provider in TypeScript or Python. An example for DB schema is shown above.

3. Indexing with RAG: we build a semantic index of the code using embeddings (CodeBERT or text-embedding-3-small) and a vector database (ChromaDB, pgvector). The index updates on repository pushes.

4. Fine-tuning (optional): we fine-tune the model on your historical PRs and typical tasks to improve suggestion relevance. We use LoRA to save resources.

As a result, the assistant suggests code that follows your conventions, not abstract examples.

Practical case: rollout to a 12-developer team

Starting state: team used GitHub Copilot, complained about irrelevant suggestions — Copilot didn't know internal patterns of a Django project with 800+ models.

Solution: Continue.dev + local Ollama for autocomplete + Claude via API for chat/refactoring + custom context provider with codebase index.

Infrastructure: server with RTX 4090 (Qwen2.5-Coder 7B for autocomplete), Claude API for complex requests.

Results after 2 months:

• Inline suggestion acceptance: 23% (Copilot) → 41% (custom) — 1.78x better than Copilot.
• Average time to write a typical CRUD endpoint: 52 min → 31 min (40% faster).
• Tasks like "write a test for this function": 100% manual → 70% automated.
• Subscription savings: over 50%, saving $2,500 per month.

Key factor for acceptance rate improvement: the context provider with codebase index gave the model real examples from the project, not abstract code.

Local models for completion

For teams with code confidentiality requirements — a fully local stack. To maximize GPU utilization we use INT4 quantization.

For inline completion, latency <200 ms is critical — users notice delay. Therefore models up to 7B are used for FIM (fill-in-the-middle).

Timelines and process

What's included

• Configuration and architecture documentation
• Access to selected models (local or cloud)
• Team training (2-hour workshop)
• Technical support and one-month warranty
• Source code for custom context providers (if developed)

Total: 3–5 weeks to full implementation. Pricing starts from $15,000 for a standard team. Contact us to get a consultation and project estimate.

We help teams of any size, from startups to enterprise with custom security requirements. We have over 5 years of experience in AI/ML and 20 implemented projects. We provide a warranty on integration and post-implementation support.

Order a custom AI assistant for your IDE — reach out, and we'll tell you in detail how to accelerate your development. Get a consultation — we'll evaluate your project.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.