justin
cd4a0f3148
CI was failing on the "Rebuild indexes from committed corpus" step
with httpx.ConnectError [Errno 111] — `localhost:11434` in the
OLLAMA_URL pool resolves to the Gitea Actions runner CONTAINER's
own localhost (no Ollama there), not the host. Fix: drop localhost
from CI's pool; it stays useful for dev runs from the workstation
where the TITAN X serves Ollama on the host loopback.
Final CI pool — 3 LAN endpoints, weighted to .0.125 (4090):
.0.125:11434 ×4 (RTX 40-series, 242 embeds/sec)
.0.2:11436 ×2 (GPU-pinned, 108 embeds/sec)
.0.2:11435 ×1 (GPU-pinned, 72 embeds/sec)
deploy/docker-compose.yml — rewrite to match Drawbar's actual
parent-stack pattern, learned by inspecting how chem-mcp is
deployed on trashpanda:
- Service name `seed-mcp` (matches chem-mcp's pattern). Reached
via docker DNS as `seed-mcp:8080` from drawbar-backend-api.
- Internal-only (no host port), expose 8080 only.
- MCP_PORT=8080 inside container (chem-mcp uses 8080 too).
- OLLAMA_URL via host.docker.internal:11434 (trashpanda's Ollama
runs on the host). extra_hosts maps host-gateway.
- RERANK_URL: http://llama-rerank:8080 — but llama-rerank is on
the default `bridge` network, not drawbar-backend_default,
so chem-mcp's reranker silently fails! Documented patch:
docker network connect drawbar-backend_default llama-rerank
Fixes rerank for BOTH chem-mcp (today: dense-only fallback)
and the new seed-mcp.
- Watchtower label set so CI pushes to :latest auto-deploy.
Documented llama-rerank service block as an alternative for
bringing the sidecar fully into the parent compose stack, with the
ubatch-size flag the seed corpus needs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
115 lines
4.7 KiB
YAML
115 lines
4.7 KiB
YAML
# seed-mcp service block to MERGE into Drawbar's parent compose file
|
|
# at /home/justin/drawbar/drawbar-backend/docker-compose.yml on
|
|
# trashpanda.
|
|
#
|
|
# Pattern matches the existing chem-mcp service (crop-chem-docs):
|
|
# internal-only, no host port, MCP_PORT=8080 inside container,
|
|
# reached via docker DNS as `seed-mcp:8080` from drawbar-backend-api.
|
|
# Ollama lives on the host (host.docker.internal); the shared
|
|
# llama-rerank sidecar must be attached to drawbar-backend_default
|
|
# (see "llama-rerank patch" section below).
|
|
|
|
services:
|
|
|
|
# seed-mcp — 760 variety identity records + 4,313 trial documents.
|
|
# Tools land in the advisor's catalog under the `seed:` prefix via
|
|
# the mcp_client multiplex (same pattern chem-mcp uses).
|
|
seed-mcp:
|
|
image: git.jpaul.io/justin/seed-mcp:latest
|
|
environment:
|
|
MCP_TRANSPORT: streamable-http
|
|
MCP_HOST: 0.0.0.0
|
|
MCP_PORT: "8080"
|
|
# FastMCP DNS-rebinding protection — disabled because we're
|
|
# only reachable on the internal docker network as
|
|
# `seed-mcp:8080`; not published to the host.
|
|
MCP_DISABLE_DNS_REBINDING_PROTECTION: "1"
|
|
PRODUCT_NAME: crop_seed
|
|
# Query-time embeddings via the host's Ollama on :11434
|
|
# (nomic-embed-text). host.docker.internal is mapped below.
|
|
OLLAMA_URL: ${SEED_OLLAMA_URL:-http://host.docker.internal:11434}
|
|
EMBED_MODEL: ${SEED_EMBED_MODEL:-nomic-embed-text}
|
|
# Reranker. Shared llama.cpp sidecar — see "llama-rerank patch"
|
|
# below for the network-attach that makes this resolvable.
|
|
RERANK_URL: ${SEED_RERANK_URL:-http://llama-rerank:8080}
|
|
RERANK_POOL: "50"
|
|
RERANK_TIMEOUT: "30"
|
|
# Hybrid + rerank is the eval-validated config (100% pass, 90%
|
|
# P@1, 0.905 MRR on 21 golden queries; see eval/results/
|
|
# baseline.md). Without rerank, P@1 drops to 62%.
|
|
HYBRID_SEARCH: "true"
|
|
RRF_K: "60"
|
|
extra_hosts:
|
|
- "host.docker.internal:host-gateway"
|
|
expose:
|
|
- "8080"
|
|
restart: unless-stopped
|
|
labels:
|
|
# Watchtower auto-pulls :latest on push from CI. The label is
|
|
# required because the Drawbar stack's watchtower is set to
|
|
# label-mode (WATCHTOWER_LABEL_ENABLE=true).
|
|
com.centurylinklabs.watchtower.enable: "true"
|
|
|
|
|
|
# ─── llama-rerank patch ──────────────────────────────────────────────
|
|
#
|
|
# As of 2026-05-25, the llama-rerank container is on the default
|
|
# Docker `bridge` network — NOT on `drawbar-backend_default` where
|
|
# chem-mcp and seed-mcp live. The chem-mcp container's
|
|
# RERANK_URL=http://llama-rerank:8080 has been silently failing to
|
|
# resolve (returns 167.100.x via public DNS, connection refused),
|
|
# falling back to dense-only retrieval. To fix this and unlock
|
|
# rerank for BOTH chem-mcp and seed-mcp:
|
|
#
|
|
# docker network connect drawbar-backend_default llama-rerank
|
|
#
|
|
# This is idempotent and survives container restarts as long as the
|
|
# llama-rerank container is recreated by Watchtower; if Watchtower
|
|
# replaces it, re-run the connect command (or — better — bring
|
|
# llama-rerank into the compose stack so the network attachment is
|
|
# declarative).
|
|
#
|
|
# Alternatively, declare llama-rerank as a compose service in this
|
|
# parent stack:
|
|
#
|
|
# llama-rerank:
|
|
# image: ghcr.io/ggml-org/llama.cpp:server-cuda
|
|
# container_name: llama-rerank
|
|
# restart: unless-stopped
|
|
# volumes:
|
|
# - llama-rerank-cache:/root/.cache/huggingface
|
|
# command: >
|
|
# -hf gpustack/jina-reranker-v2-base-multilingual-GGUF:Q8_0
|
|
# --reranking
|
|
# --host 0.0.0.0 --port 8080
|
|
# --n-gpu-layers 99 --ctx-size 8192
|
|
# --batch-size 4096 --ubatch-size 4096 --parallel 4
|
|
# deploy:
|
|
# resources:
|
|
# reservations:
|
|
# devices:
|
|
# - driver: nvidia
|
|
# count: 1
|
|
# capabilities: [gpu]
|
|
# labels:
|
|
# com.centurylinklabs.watchtower.enable: "false"
|
|
# expose:
|
|
# - "8080"
|
|
#
|
|
# (Watchtower disabled on llama-rerank to avoid surprise model-reload
|
|
# downtime; pin the image tag if you want predictability.)
|
|
#
|
|
# Note: --ubatch-size 4096 is required for the seed-mcp corpus —
|
|
# default 512 is too small for the ~600-token trial chunks and
|
|
# rejects the whole batch with "input too large to process".
|
|
|
|
# ─── drawbar-backend-api wiring ──────────────────────────────────────
|
|
#
|
|
# Add to the drawbar-backend-api `environment:` block so it can call
|
|
# seed-mcp alongside chem-mcp:
|
|
#
|
|
# SEED_MCP_BASE_URL: ${SEED_MCP_BASE_URL:-http://seed-mcp:8080/mcp}
|
|
#
|
|
# The advisor's tool multiplex (mcp_client) picks this up and exposes
|
|
# the seed-mcp tools under the `seed:` prefix.
|