Skip to content

code-intelligence.md

Latest commit

 

History

History
143 lines (109 loc) · 5.8 KB

code-intelligence.md

File metadata and controls

143 lines (109 loc) · 5.8 KB
Raw

Code Intelligence

Local AI Stack can use separate local models for general chat, code generation, and low-latency inline completion. The code-intelligence tools never upload a repository to a cloud service.

Install code mode

.\scripts\setup.ps1 -CodeMode
# or: .\scripts\pull-models.ps1 -CodeMode
bash scripts/setup.sh --code-mode
# or: bash scripts/pull-models.sh --code-mode

This pulls qwen2.5-coder:7b for code generation and bug fixing, plus qwen2.5-coder:1.5b for inline completion. The OpenCode config exposes both models; select the 7B model for edits and the 1.5B model for short completions.

FastGPT routing

FastGPT workflows are stored in a user's FastGPT database, so this repository does not overwrite them. Add a Condition node before the chat model node:

  1. Route to qwen2.5-coder:7b when the input contains a code fence, an error trace, or terms such as bug, function, 写代码, 修复 bug.
  2. Route all other prompts to qwen3:8b.
  3. For editor completion clients, use qwen2.5-coder:1.5b.

code_intelligence/router.py contains the deterministic rules and can be used by a custom FastGPT HTTP/tool node without calling a model to make this choice.

AST-aware code indexing

python code_intelligence/index_codebase.py /path/to/project --output code-index.json

The JSON output contains function/class/method chunks, signatures, docstrings, per-file imports, called symbol names, and Python/JavaScript import edges. Use these chunks as the source for a code-specific embedding model or import them into a separate FastGPT code knowledge base. Keeping code and prose knowledge bases separate lets each use an appropriate embedding model.

For local CodeBERT embeddings, install the optional dependencies and embed the index. Code chunks receive the code embedding profile while ordinary prose can remain on nomic-embed-text:

pip install -r code_intelligence/requirements-code-embeddings.txt
python code_intelligence/embed_chunks.py code-index.json --output code-embeddings.json

To include this index in the scheduled knowledge-base sync, set CODE_INDEX_ENABLED=true and LOCAL_AI_STACK_ROOT in knowledge-base/sync/config.env. Optionally set CODE_INDEX_ROOT to a specific repository; otherwise the knowledge-base root is indexed.

Safe execution and project operations

Run only user-approved commands with an isolated container:

python code_intelligence/sandbox.py --workspace /path/to/project \
  --image python:3.12-alpine --command "python -m pytest"

The sandbox has no network, a read-only project mount, no Linux capabilities, PID/CPU/memory limits, and a short timeout. It cannot apply patches or run Git operations. Keep project mutations outside the sandbox and present the exact git command or patch to the user for explicit confirmation.

code_intelligence/project_actions.py provides two confirmation-gated local operations (commit and create-file). It always previews a mutation first; only the identical command with --confirm performs it.

LSP and feedback

Start the lightweight local completion server with:

python code_intelligence/lsp_server.py

Point an LSP-capable editor at this stdio command. It implements textDocument/completion and calls only the local Ollama API. A VS Code client can use the same command through its LSP extension settings. The server keeps a small in-memory LRU cache for identical recent prefix requests; it never writes editor content to disk.

The included ide/vscode extension provides inline completion and Local AI: Explain Selection / Suggest Fix for Selection using the local Ollama API. Fixes are displayed as an unapplied diff; it never applies an edit or executes a command on behalf of the user. Its chat answer includes thumbs-up and thumbs-down buttons. A click is the explicit approval to append that answer to the workspace-local, git-ignored lora-finetune/data/feedback.jsonl (configurable with localAiStack.feedbackPath); feedback still requires human review before LoRA training.

Store fine-tuning feedback only after explicit approval:

python code_intelligence/feedback.py --approved --rating down \
  --prompt "..." --response "..." --correction "..."

Approved feedback is written to the git-ignored lora-finetune/data/feedback.jsonl; review and redact it before converting it to a training dataset.

The Monthly Approved Feedback LoRA workflow runs on the first day of each month on a self-hosted GPU runner. Set its LOCAL_AI_FEEDBACK_FILE and LOCAL_AI_APPROVED_DATA_FILE repository variables to absolute runner paths, or provide both paths when dispatching it manually. It converts candidates but refuses to train until the approved data file exists.

Self-hosted quality gate

The Local Code Quality Evaluation workflow is manually dispatched on a self-hosted runner labeled ollama. It runs a local benchmark and can compare against a baseline JSON. The default regression limit is 3%; a larger decline returns a failing workflow. Replace benchmarks/codegen-smoke.json with a licensed HumanEval/MBPP-compatible dataset for a broader release gate.

On Windows, register the local machine once (requires an authenticated gh CLI and Ollama) and then keep the runner process open:

.\scripts\setup-actions-runner.ps1
& 'C:\local-ai-stack-runner\run.cmd'
.\scripts\install-runner-startup.ps1

The setup script obtains a short-lived registration token without printing or persisting it, and registers labels ollama,gpu. After the runner appears online in the repository's Settings → Actions → Runners, dispatch Local Code Quality Evaluation from the Actions tab. The startup helper creates a current-user shortcut, so the runner starts after Windows sign-in without requiring administrator permission.