ai-code-intelligence-claude-context

A stand-alone reproducible lab for @zilliz/claude-context against the AI-Code-Intelligence research repo liatrio/ai-code-intelligence, tracking issue #84.

Sibling PoCs (same shape, different tool): ai-code-intelligence-semgrep, ai-code-intelligence-codebase-memory, ai-code-intelligence-socraticode, ai-code-intelligence-gitnexus, ai-code-intelligence-graphify, ai-code-intelligence-repowise.

Family

Indexing. claude-context is a semantic-retrieval MCP server: AST-aware chunks embedded into a vector index, hybrid BM25 + dense-vector search, Merkle-tree incremental sync. Answers "where does the code for concept X live?" and "is there already a helper that does Y?" It does not answer "who calls X?" — that is a structural-graph tool's job.

What this lab measures

Six-wave protocol against the fully-local claude-context path (Milvus in Docker Compose + Ollama for embeddings). The lab deliberately does not exercise the quickstart-default hosted path (Zilliz Cloud + OpenAI); see AGENTS.md for why, and the no_default_egress cell on research/claude-context/data.json for what the lsof probe covers when the two paths are compared.

• Fixtures: liatrio/gratibot (public, JS/Node, 195 files) and liatrio-labs/liatrio-knowledge (private, TS/SQL/Gherkin/Docusaurus, 5,370 tracked files).
• Harness arms: baseline (Read/Grep only) vs claude-context (Read/Grep/Bash plus the pinned MCP server).
• Prompts: five semantic-retrieval prompts with pass conditions frozen in prompts.md before any run.
• Waves: install/index smoke, yes-cell confirmations, no-cell confirmations, shared/gateway probes, gratibot harness, liatrio- knowledge harness. Results in RESULTS.md.

Prerequisites

• Node ≥ 20, npm ≥ 10 (claude-context's monorepo engines minimum).
• Docker ≥ 20.10.13 with docker compose v2 (for Milvus standalone).
• Ollama ≥ 0.3.0, running on 127.0.0.1:11434.
• Claude Code CLI (claude or claude-code) on PATH.

No cloud accounts, no API keys, no login flow. setup.py --check will refuse to proceed if any hosted-provider env var is set (see FORBIDDEN_ENV_VARS in versions.env).

Contract files (#75)

Contract file	Purpose
setup.py	--check, --install, --index, --status, --clean
AGENTS.snippet.md	Paste-in snippet for a client project's AGENTS.md
prompts.md	5 prompts + frozen pass conditions
RESULTS.md	Wave-by-wave findings; grows as waves land
run-prompts.py	Two-arm Claude Code harness

Usage

Clone a fixture outside this repo:

git clone https://github.com/liatrio/gratibot.git ~/liatrio/repos/gratibot

Bring up the local stack and index the fixture:

make check                                                # env gate
make install                                              # npm + docker compose + ollama pull
python3 setup.py --index --fixture ~/liatrio/repos/gratibot

Run the harness:

python3 run-prompts.py --fixture ~/liatrio/repos/gratibot

Tear it all down:

make clean

Where the numbers land

• Wave-by-wave rationale and probe details: RESULTS.md.
• Cells settled by the lab: research/claude-context/data.json.
• Cell rationale: research/claude-context/answers.md.

Licence

The lab code in this repository is MIT. @zilliz/claude-context-mcp and @zilliz/claude-context-core are MIT. Milvus is Apache-2.0. Ollama is MIT. nomic-embed-text (the pinned embedding model) is Apache-2.0.