main
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 justin
|
cd4a0f3148 |
CI fix + Drawbar-stack deploy pattern
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>
|
||
|
justin
|
0e625553e5 |
gh_plot_reports corpus (4,299 plots) + concurrency + 4-GPU pool
CORPUS — 4,299 GH plot reports added (3,797 written + 502 from the
earlier slow run + 319 sitemap-listed URLs that 404'd as
discontinued). Combined with prior 760 varieties + 14 AgriPro
trials = 5,073 total chunks now indexed.
scrape/sources/gh_plot_reports.py — concurrency speedup:
- 4 worker threads (ThreadPoolExecutor), each with its own
requests.Session for connection-pool efficiency.
- Shared class-level rate limiter (0.25 sec between ANY two
requests across all threads). Net throughput ~4 req/sec —
well below any rate-limit threshold a public site enforces.
- Diagnosis vs original 1 req/sec: GH had ZERO rate limiting,
zero 429s, zero retries. The 1 sec self-throttle was just too
conservative. Bench:
1 worker / 1.0 sec throttle: ~0.4 plots/sec (190 min ETA)
4 workers / 0.25 sec throttle: ~3 plots/sec (~25 min actual)
rag/chunk.py — chunk size cap for nomic-embed-text's 2048-token
context window:
- Empirically tested: failure threshold is ~5,250 chars on
numeric-heavy trial chunks (chars/token ratio 2.4 vs 3.5 for
prose). Cap at 4,500 chars to be safely under at worst-case
2.2 chars/token.
- Applied to BOTH variety and trial chunks. Marked truncated
chunks with metadata.embed_truncated = True; FULL text stays
in the on-disk .md for get_page to return verbatim.
.gitea/workflows/{refresh,image-only}.yml — OLLAMA_URL pool
restructured for the 4 GPU-pinned endpoints. Bench (50-chunk
batches on nomic-embed-text):
.0.125:11434 (RTX 40-series) 242 embeds/sec ← weight ×4
.0.2:11436 (GPU-pinned) 108 embeds/sec ← weight ×2
.0.2:11435 (GPU-pinned) 72 embeds/sec ← weight ×1
localhost (TITAN X) 37 embeds/sec ← weight ×1
Weighting is done by listing the URL multiple times in
OLLAMA_URL since the embedder uses round-robin. .0.2:11434 is
explicitly EXCLUDED — it isn't pinned to a specific GPU.
Combined index rebuild for 5,073 chunks now finishes in ~3 min
(was 19+ on the single-endpoint pool).
Smoke tests:
✓ list_versions: 5,073 docs across 6 sources, 2 vendors, 6
brands, 4 crops (corn 2711, soy 2016, silage 223, wheat 123).
✓ search_trials({crop=corn, state=IA, year=2024}): 3 IA 2024
corn trials surfaced.
✓ search_trials("Phytophthora resistance soybean trial"): NK
NK43-W1XFS top-1 in LA 2024 trial (cross-vendor result).
✓ search_trials("AP Iliad Idaho wheat"): AgriPro Washington/N
Idaho 2025 trial surfaced.
✓ search_trials(product=DKC65-95): 3 corn trials containing
that hybrid in IL/IA 2024.
✓ search_trials(product=NK1701): 3 corn trials in AR/MS 2024.
✓ Product filter correctly returns EMPTY for products that
aren't in the corpus (DKC65-20 is a 2023 product; 2023 plots
deferred). Anti-hallucination contract preserved.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
justin
|
ac40e05734 |
seed-mcp scaffold: clone docs-mcp-template, customize for crop_seed PRODUCT_NAME
Image rebuild (skip scrape) / build (push) Failing after 7s
Sibling project to crop-chem-docs, same MCP-template lineage. Corpus is
seed/hybrid varieties across 6 vendors instead of pesticide labels.
What's customized vs. the template:
- CLAUDE.md: vendor matrix, build priority, Pioneer fallback policy,
canonical sidecar schema (per-crop), Golden Harvest disease-scale
reversal gotcha, no-IPv6 / HTTPS-clone note
- README.md: vendor coverage table, tool list, phase status
- Dockerfile: PRODUCT_NAME=crop_seed default, sources.json (not
bundles.json), HYBRID_SEARCH=true, OLLAMA_URL + RERANK_URL Docker
DNS defaults (same llama-rerank sidecar as crop-chem-docs)
- .gitea/workflows/refresh.yml: monthly cron (seed catalogs move
slowly), 5 GREEN scraper steps, corpus-YYYY.MM.DD tag for Drawbar
pinning, continue-on-error on GC step
- .gitea/workflows/image-only.yml: paths filter + cancel-in-progress
concurrency group
- scripts/registry_gc.py: lifted from crop-chem-docs (correct Gitea
packages API URL + UA header to bypass CF block on default
Python-urllib UA)
- sources.json: catalog of 6 sources + scope_filter + per-source
schema notes + Pioneer-exclusion rationale
- scrape/runner.py: dispatcher with --all = GREEN-only
- scrape/sources/{bayer_seeds,golden_harvest,nk,agripro,becks_pfr,
becks_products}.py: stub modules with implementation notes
- docs_mcp/server.py: PRODUCT_NAME default → crop_seed,
PRODUCT_DOCS_URL → repo URL
Pioneer is intentionally NOT a source. ToS bans automation; dealer
locator is login-gated. The MCP returns a curated fallback lesson
directing the user to pioneer.com.
Next phases:
- Phase 1: implement bayer_seeds (lift-and-shift from crop-chem-docs
Bayer scraper; same __NEXT_DATA__ infra)
- Phase 7: curate eval/queries.jsonl
- Phase 11: lessons.md with Pioneer fallback + disease-scale notes
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|