justin
bd71f30ca7
Phase 6 — Reranker integration
- New _rerank(query, [(cid, doc), ...]) helper in server.py calls
llama.cpp's /v1/rerank endpoint, returns reranker-ordered ids
or None on failure (graceful fallback — search never blocks
on the sidecar).
- search_docs + search_trials both call _rerank() on the post-
hybrid pool BEFORE truncating to k. The variety-code prefilter
still pins exact matches on top.
- Per-doc truncation to 2000 chars to fit jina-reranker-v2-base's
per-pair token budget. Full chunk text still returned to the
caller — truncation is rerank-input-only.
- Telemetry adds `reranked: true|false` so usage logs distinguish
reranked calls.
Phase 7 — Eval harness
- eval/queries.jsonl: 21 golden queries spanning:
* variety-code lookups (DKC62-08RIB, AG29XF4, WB6430, E085Z5,
AP Iliad)
* semantic variety queries (drought-tolerant corn, SCN MG-3
soy, Rps3a, XtendFlex, HRS stripe rust, SWW PNW, Goss's Wilt)
* trial queries (IA/IN/MN regional, AP Iliad ID, NK1701 head-
to-head, silage Ton/Acre, product=DKC65-95)
* anti-hallucination (Pioneer P1142 fallback, DKC65-20 not-in-
corpus expected_empty)
- eval/retrievers.py: 4 named retrievers — dense, bm25, hybrid
(dense+bm25+RRF), hybrid+rerank — all sharing the same filter
shape as docs_mcp/server.py._build_where.
- eval/run_eval.py: runs each retriever against each query,
reports Recall / Precision@1 / MRR / avg latency. Markdown
output in eval/results/baseline.md.
Baseline results (k=5, 21 queries):
| Retriever | Pass | Recall | P@1 | MRR | Avg ms |
|-----------------|-------|--------|-------|-------|--------|
| hybrid+rerank | 21/21 | 100% | 90% | 0.905 | 2064 |
| bm25 | 20/21 | 95% | 81% | 0.833 | 5 |
| hybrid | 15/21 | 71% | 62% | 0.619 | 73 |
| dense | 14/21 | 67% | 38% | 0.440 | 79 |
Key findings:
1. hybrid+rerank wins on quality — 100% pass, 90% P@1.
2. BM25 alone is surprisingly competitive (95% pass) at 5 ms —
excellent fallback when rerank is down. The variety-code
prefilter in search_docs is doing a lot of work here.
3. Dense embedding alone is the WEAKEST configuration on this
corpus — variety identity tokens (DKC62-08RIB, AP Iliad,
Rps3a) have no semantic neighbors, so nomic-embed-text returns
noise. The hybrid (no rerank) layer actively hurts because
RRF dilutes the BM25 ranking with dense noise.
4. Anti-hallucination queries (Pioneer fallback, DKC65-20 not-
in-corpus) pass on ALL retrievers including dense-only —
the must_not_contain + expected_empty design holds.
Deploy decision: HYBRID_SEARCH=true + RERANK_URL set
(production env already has both — refresh.yml + image-only.yml
+ deploy/docker-compose.yml all configured).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
seed-mcp
MCP server over the public catalogs of major US row-crop seed
vendors — corn, soybeans, wheat. Sibling project to
crop-chem-docs
(pesticide labels), feeding the same Drawbar farm-advisor AI.
The server exposes per-variety records with agronomic ratings, disease tolerance, trait stack, maturity, and regional notes — so the advisor can answer questions like "which corn hybrid for sandy soil, drought-prone, RM ≤105 in northeast Iowa?" without rummaging through individual brand sites.
Vendor coverage
| Vendor | Verdict | Varieties | Notes |
|---|---|---|---|
| Bayer seeds (DEKALB + Asgrow + WestBred) | 🟢 | ~475 | Same cropscience.bayer.us Next.js infra as crop-chem-docs |
| Golden Harvest (Syngenta) | 🟢 | ~175 | Sitemap + server-rendered HTML + Syngenta CDN PDFs |
| NK (Syngenta) | 🟢 | 29 | Shares PDF fetcher with Golden Harvest |
| AgriPro (Syngenta wheat) | 🟢 | 24 | Drupal Views, server-rendered |
| Beck's PFR | 🟡 | 2,089 | Public Sanity GROQ API (no auth) |
| Beck's products | 🟡 | 860 | Identity-only until SeedIQ XHR sniffed |
| Pioneer (Corteva) | 🔴 | — | ToS bans automation — curated fallback lesson instead |
Quick start
git clone https://git.jpaul.io/justin/seed-mcp.git
cd seed-mcp
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# Run one scraper
python -m scrape.runner --source bayer_seeds --force
# Rebuild indexes
python -m rag.index --rebuild
# Local MCP server (stdio for Claude Desktop dev)
python -m docs_mcp.server --transport stdio
Tools exposed
| Tool | Purpose |
|---|---|
search_docs |
Hybrid + rerank variety search with crop / RM / trait / region filters |
get_page |
Full variety record by (source, source_key) |
list_versions |
Discover crops, brands, traits, RM/MG ranges, wheat classes |
corpus_status |
Counts + freshness; useful for health probes |
crop_seed_api_lessons |
Curated agronomy lessons — Pioneer fallback, disease-scale normalization, regional placement heuristics |
Build phases
This is a clone of docs-mcp-template.
The 13 phases in PLAN.md apply:
| Phase | Status |
|---|---|
| 0 — scaffold | done |
| 1 — first scraper (bayer_seeds) | next |
| 2 — chunk + index | pending |
| 3 — baseline MCP tools | template defaults |
| 4-5 — Dockerfile + CI | done (placeholders filled) |
| 6 — reranker | shares llama-rerank sidecar with crop-chem-docs |
| 7 — eval harness | pending (curate ~25 queries) |
| 8 — hybrid search | done (template) |
| 9 — diff_versions, list_cluster | optional |
11 — crop_seed_api_lessons curated layer |
pending |
See CLAUDE.md for the canonical sidecar schema and the
disease-scale-normalization gotcha (Golden Harvest is reversed).
Infrastructure
- Registry:
git.jpaul.io/justin/seed-mcp:latest(Watchtower) /:corpus-YYYY.MM.DD(production pin) - Embedder: shared Ollama pool with crop-chem-docs (Gitea-host GPUs + Windows Ollama; CI never hits trashpanda's production Ollama)
- Reranker: shared
llama-reranksidecar on trashpanda's Tesla P4 (one container, both MCPs use it) - PRODUCT_NAME:
crop_seed(notseed_mcp— used in Chroma collection, BM25 db filename, andcrop_seed_api_lessonstool)