Ollama MCP Langgraph Lab

ollama-mcp-langgraph-lab

Docker Compose lab‑stack: Ollama + MCP (FastMCP) + LangChain/LangGraph агент. Всё настраивается через .env.

Overview Four services in one Compose:

1. Ollama LLM server (pulls and runs models).
2. FastMCP server exposes tools over HTTP at /mcp.
3. L0 FastMCP server exposes deterministic SPIR tools at /mcp.
4. LangChain/LangGraph agent CLI that uses ChatOllama + MCP tools.

Requirements

1. Docker + Docker Compose.
2. For GPU mode: NVIDIA driver + nvidia-container-toolkit (Linux/WSL2). macOS has no GPU passthrough.

Quickstart

cp .env.example .env
mkdir -p workspace

# default is GPU now; use "cpu" explicitly if needed
./lab.sh up
./lab.sh pull qwen3:8b
./lab.sh ask "Сделай файл report.txt с текущим временем UTC и перечисли файлы в workspace"

Common Commands

./lab.sh up [gpu|cpu]
./lab.sh down [gpu|cpu]
./lab.sh pull 
./lab.sh list
./lab.sh ask "your prompt"
./lab.sh logs

BDD + Allure

behave -f allure_behave.formatter:AllureFormatter -o allure-results
allure serve allure-results

Default testcontainers compose file: tests/compose/docker-compose.test.yml (L0 + hydra_mock).

Important Notes

1. Default profile is gpu (see lab.sh). Use ./lab.sh up cpu if you do not have GPU support.
2. Ollama is exposed on http://localhost:11434.
3. MCP server is exposed on http://localhost:8000/mcp.
4. L0 MCP server is exposed on http://localhost:8001/mcp.
5. Model, context size, and reasoning mode are controlled in .env.
6. Flow mode runs the explicit graph: `python /app/flow.py "

"`.

L0 Syntax + Semantics Layer (SPIR v0.5.0) L0 now produces a structured syntax+semantics package:

1. syntax.paninian_edges: list of {head, dep, role} (karaka graph).
2. syntax.ud.basic_edges: basic UD tree (rel labels, exactly one root).
3. syntax.ud.enhanced_edges + syntax.ud.empty_nodes: UD-compatible enhanced layer (i.j empty nodes + DEPS).
4. syntax.clauses + syntax.discourse_links: clause/discourse layer.
5. semantics.kag: Event+Deontic KAG with provenance.
6. Clause spans use half-open policy: token_span = [start, end).

Backend selection:

1. SYNTAX_BACKEND=rules (default) uses internal rule-based attachment.
2. SYNTAX_BACKEND=hydra / hyderabad uses the external Hyderabad/Samsaadhanii parser.
3. SYNTAX_BACKEND=none disables syntax.

UD mode:

1. ud_mode=head_rules (default) uses runtime head_rules.yaml.
2. ud_mode=projected uses direct Paninian->UD projection (debug/compat inside v0.5 only).
3. ud_mode=none disables UD.

Hyderabad parser wiring:

1. HYD_PARSER_URL to call a running HTTP/CGI service.
2. HYD_PARSER_CMD to execute a local CLI command (use {text} placeholder or stdin).
3. KARAKA_UD_MAP_PATH to override karaka->UD mapping JSON.
4. UD_HEAD_RULES_PATH to override active UD head rules YAML.
5. SYNTAX_OVERRIDES_PATH to apply deterministic per-input graph patches.
6. RETRIEVAL_BACKEND=baseline|hybrid_prod to choose retrieval tier.

New L0 MCP tools:

1. l0_analyze (SPIR v0.5 output).
2. l0_query_understand (NL|SPIR -> KAG query + retrieval plan).
3. l0_retrieve (baseline|hybrid_prod tiered retrieval with fallback metadata).
4. l0_export (sidecars: CoNLL-U basic/enhanced, KAG JSONL, align JSON).

CLI example:

SYNTAX_BACKEND=rules python -m sprs_l0.cli analyze --in src/l0/data/prepared/corpus.jsonl --out workspace/spir_samples_v05.jsonl
python -m sprs_l0.cli validate --in workspace/spir_samples_v05.jsonl
python -m sprs_l0.cli export --in workspace/spir_samples_v05.jsonl --out-dir workspace/exports