dc-research-mcp

Hybrid-retrieval MCP server for the DC Greenfield Tashkent research corpus. .NET 10 + PostgreSQL 18 + pgvector + vchord_bm25 + ONNX Runtime (CPU).

Architectural analogue of dantte-lp/arista-mcp and dantte-lp/nutanix-mcp. For a lightweight pure-Python BM25-only alternative (laptop bootstrap, no Docker) see dantte-lp/dc-research-mcp-py.

What's indexed

The corpus lives in a separate repository dc_greenfield_research. This MCP indexes:

Source	Origin
30-findings/**/*.md (research findings + follow-up reports)	research
60-tech-spec/**/*.md (Техническое задание / ТЗ)	tech-spec
00-overview/**/*.md (project overview)	overview
assets/02-standards/converted/**/*.md (TIA-942-C, ASHRAE, NFPA, BICSI, EN 50600, ShNK/KMK …)	standard
assets/01-from-uzum-tz/converted/**/*.md (vendor / former-vendor sources)	vendor

Each chunk gets metadata: source_file, origin, category (A-strategy…J-ai-specific, tech-spec, intl-standards, uz-standards), h1/h2/h3, chunk_index, llm_source (claude/openai), prompt_id (A1, C-FU-2, …).

Architecture

30-findings + 60-tech-spec + 00-overview + assets/02-standards
        │
        ▼  CLI: ingest → chunker → ONNX embedder (CPU)
   ┌────────────────────────────────────────┐
   │ PostgreSQL 18 (podman / docker)         │
   │   vector + vchord + vchord_bm25         │
   │   pg_tokenizer (russian, english, mixed)│
   │ ┌──────────────┐  ┌───────────────────┐ │
   │ │ chunks(text) │  │ embeddings(vector)│ │
   │ │ + bm25       │  │ + hnsw            │ │
   │ └──────────────┘  └───────────────────┘ │
   └────────────────────────────────────────┘
        │
        ▼  hybrid search: vchord_bm25 ⊕ pgvector HNSW → RRF
   MCP Server (stdio) — ModelContextProtocol 1.2
        │
        ▼
   Claude Code / any MCP client

Quick start

Prerequisites: .NET 10 SDK (10.0.201 or newer), Podman 5+ (or Docker), and access to the dc_greenfield_research corpus on disk.

# 1. start Postgres with vchord + vchord_bm25 + pg_tokenizer
podman compose -f docker/compose.yaml up -d
# (Postgres listens on host port 5435; arista-mcp uses 5434 to avoid collision)

# 2. restore + build
dotnet restore
dotnet build -c Release

# 3. run migrations (sprint 1)
# dotnet run --project src/DcResearchMcp.Cli -- migrate

# 4. ingest the corpus (sprint 3)
# DC_RESEARCH_SOURCE_DIR=C:/SHARE/dc_greenfield_research \
#   dotnet run --project src/DcResearchMcp.Cli -- ingest

# 5. start MCP server (sprint 4) — stdio
# dotnet run --project src/DcResearchMcp.Server

Connecting to Claude Code

~/.claude/mcp.json or per-project .mcp.json:

{
  "mcpServers": {
    "dc-research": {
      "command": "dotnet",
      "args": [
        "run",
        "--project",
        "C:/SHARE/dc-research-mcp/src/DcResearchMcp.Server",
        "-c", "Release",
        "--no-build"
      ],
      "env": {
        "DC_RESEARCH_SOURCE_DIR": "C:/SHARE/dc_greenfield_research",
        "DC_RESEARCH_PG_CONN": "Host=localhost;Port=5435;Database=dc_research;Username=dc_research;Password=dc_research"
      }
    }
  }
}

Variants

Use case	Project
Laptop bootstrap, no Docker, no DB, ~50 MB RAM	dc-research-mcp-py (BM25-only, bm25s)
Workstation, Docker available, hybrid retrieval (this repo)	dc-research-mcp (.NET 10 + Postgres + ONNX CPU)
Pattern for other corpora	nutanix-mcp, arista-mcp

Status

Sprint 0 (scaffolding) — initial release 0.1.0. Solution + 5 projects + 5 test projects + Postgres compose with vchord_bm25 + analyzers (russian, english, mixed). dotnet build passes on empty stubs; podman compose up brings up Postgres with required extensions. Functional code lands in subsequent sprints — see CHANGELOG.md.

Roadmap (planned):

• Sprint 1 — Core + Data (models, EF Core, migrations, COPY loader)
• Sprint 2 — Embedding (ONNX wrapper + multilingual-e5-small)
• Sprint 3 — Cli (ingest / reindex / search / info)
• Sprint 4 — Server (MCP tools: search, get_chunk, list_sources, list_files, index_info)
• Sprint 5 — Testcontainers + xunit + GitHub Actions CI
• Sprint 6 — 0.2.0 release

License

MIT. See LICENSE.