Phase 2/3: chunker + indexer + MCP server tools

Phase 2 — Chunking and indexing
- rag/chunk.py: replace template chunker with seed-variety-specific
  chunks_from_variety(). One chunk per variety (varieties are small
  and named-rating retrieval signal is best kept together). Output
  is rebuilt deterministically from the sidecar JSON: every value is
  verbatim from the source, only framing language ("Disease ratings
  (1-9, 9=best):") is template glue. Anti-hallucination contract:
  same sidecar in → same chunk out, never a fabricated rating.
  Metadata flattened to Chroma-safe primitives (str/int/float/bool):
  source, source_key, vendor, brand, crop, product_name,
  product_id, source_url, rm (corn), mg (soy), wheat_class,
  release_year, trait_codes_csv, rating_scale.
- rag/index.py: walks corpus/<source>/<source_key>.json sidecars
  via the new chunker. Default PRODUCT_NAME=crop_seed so the
  Chroma collection is crop_seed_docs.
- rag/bm25.py: filterable columns updated to seed-domain facets
  (source/vendor/brand/crop/source_key) instead of the template's
  version/platform/product.

Phase 3 — MCP server tools wired up
- search_docs: hybrid dense (Chroma) + BM25 (FTS5) retrieval with
  RRF fusion. Optional filters: crop, brand, vendor, source.
  Variety-code prefilter pins exact source_key / product_name /
  hybrid_prefix matches at the top — dense embeddings have no
  semantic neighbor for tokens like "DKC62-08RIB" and RRF can let
  noise float to #1 without this pin. Each response carries the
  variety's source URL inline so the agent can cite.
- get_page(source, source_key): emits a structured ratings header
  (verbatim from sidecar, table per characteristics group, vendor
  positioning, regional listings) followed by the raw indexed body.
  This is the canonical fact-check surface.
- list_versions(): facet discovery — distinct sources, vendors,
  brands, crops across the corpus.
- lookup_variety(source_key, source?): returns the raw sidecar JSON
  for one variety. The agent should call this BEFORE quoting any
  specific rating value to a farmer — guaranteed verbatim.

Smoke tests against 475 indexed Bayer varieties:
- list_versions returns 475 varieties, 1 source, 1 vendor, 3 brands,
  3 crops with correct per-brand counts (288/102/85).
- Semantic ag queries find the right candidates: short-season
  drought-tolerant corn → DKC44-97RIB at RM 94 (in 90-95 band);
  SCN+MG3 soybean → Asgrow XF varieties with explicit SCN R3 ratings;
  Phytophthora Rps3a soy → AG07XF4 (right gene); stripe-rust
  wheat → WestBred WB1376CLP (Yellow Rust 2 = best).
- Variety-code lookups work via prefilter: DKC62-08RIB, AG29XF4,
  WB6430 all return as #1 hit. BM25 confirms ranking unambiguously
  (top-1 score -13.2 vs -8.5 for #2 on "DKC62-08RIB ratings").
- Server boots cleanly in stdio AND streamable-http modes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-25 13:14:16 -04:00
parent 0fb8d9d92d
commit a766756a05
4 changed files with 982 additions and 369 deletions
+52 -99
View File
@@ -1,17 +1,14 @@
"""SQLite FTS5-backed BM25 retrieval over the same chunks Chroma indexes.
Hybrid retrieval (BM25 + dense + Reciprocal Rank Fusion) addresses a
limit of single-tower dense embeddings: when a query has specific
technical terms (filenames, language names, error codes, API paths),
the dense embedding doesn't bridge from the query into a short
code-focused chunk. The chunk loses to the much larger crowd of
prose chunks that semantically match the query topic.
BM25 handles this directly. Lexical overlap on rare terms ("python",
"create_vpg.py", "PROTECTED_SITE_ID", "applyUpgrade") scores those
chunks high. Fused with the dense ranking via RRF, the hybrid result
is strictly better than either alone for the queries we've seen
fail.
single-tower dense embedding's weakness on rare technical tokens —
for seed-mcp that's variety codes ("DKC62-08RIB"), trait codes
("XF", "VT2PRIB"), disease abbreviations ("SCN", "SDS", "Goss's"),
and Rps gene names ("Rps1c", "Rps3a"). Dense embeddings don't bridge
queries like "Rps3a soybean" cleanly into the relevant chunk; BM25
matches them directly. Fused with the dense ranking via RRF, the
hybrid result is strictly better than either alone for the queries
we expect from the farm-advisor agent.
Why SQLite FTS5:
- In the stdlib. Zero new deps.
@@ -19,36 +16,13 @@ Why SQLite FTS5:
`rag.index --rebuild` regenerates from corpus.
- Built-in `bm25()` ranking function. No knobs to tune that matter
for our use case (k1=1.2, b=0.75 defaults are fine).
- Builds 70k+ chunks in seconds. Faster than the Chroma rebuild's
embedding step by 100×, so it adds basically nothing to the
full-rebuild cycle.
- Builds <1k chunks in milliseconds; adds nothing to rebuild time.
Schema is two tables to keep filtering clean. FTS5 doesn't filter
nicely on its own columns; the content_rowid pattern keeps an
external metadata table joinable by rowid:
CREATE TABLE chunks_meta (
rowid INTEGER PRIMARY KEY AUTOINCREMENT,
id TEXT UNIQUE,
bundle_id TEXT, page_id TEXT, version TEXT,
platform TEXT, product TEXT, ordinal INTEGER
);
CREATE VIRTUAL TABLE chunks_fts USING fts5(
text,
tokenize = 'porter unicode61 remove_diacritics 2',
content = 'chunks_meta',
content_rowid = 'rowid'
);
Queries:
SELECT m.id, bm25(chunks_fts) AS score
FROM chunks_meta m
JOIN chunks_fts f ON m.rowid = f.rowid
WHERE f MATCH ?
AND m.version = ? -- optional metadata filter
ORDER BY bm25(chunks_fts) -- lower = better in FTS5
LIMIT ?;
external metadata table joinable by rowid. For seed-mcp the
filterable columns are seed-domain facets — source, vendor, brand,
crop, source_key — rather than the docs-template version/platform.
"""
from __future__ import annotations
@@ -63,42 +37,30 @@ log = logging.getLogger(__name__)
# Default location: bm25/<product>_docs.db at the repo root, next to chroma/.
ROOT = Path(__file__).resolve().parent.parent
DEFAULT_DB_DIR = ROOT / "bm25"
DEFAULT_DB_NAME = "<product>_docs.db"
DEFAULT_DB_NAME = "crop_seed_docs.db"
# Columns we expose as filterable metadata. Mirrors what _build_where in
# docs_mcp/server.py accepts so the same filter dicts work for both
# Chroma and BM25 without per-retriever translation in the caller.
FILTER_COLUMNS = ("bundle_id", "page_id", "version", "platform", "product", "ordinal")
# Columns we expose as filterable metadata. Mirrors what
# ``docs_mcp.server._build_where`` accepts so the same filter dict
# works for both Chroma and BM25 without per-retriever translation.
FILTER_COLUMNS = ("source", "vendor", "brand", "crop", "source_key", "ordinal")
# Allowlist tokenizer for free-text queries. FTS5's parser chokes on lots
# of punctuation we routinely see in user queries (".10.9", "?", "VPG's",
# em-dash, etc.). Rather than blocklist every operator, just keep
# alphanumerics + a few separators and replace everything else with a
# space. This loses the ability to phrase-search ("exact match") but we
# don't expose that to users anyway — they ask natural-language questions
# and want the answer, not a Boolean DSL.
# Allowlist tokenizer for free-text queries. FTS5's parser chokes on
# lots of punctuation we routinely see in farmer queries ("Rps1c",
# "SCN-resistant", "0.05 MG", em-dashes). Rather than blocklist every
# operator, keep alphanumerics + a few separators and replace
# everything else with a space.
_KEEP_RE = re.compile(r"[^A-Za-z0-9_\s]")
# FTS5 reserves these Boolean operator KEYWORDS at the token level
# stripping them avoids accidental phrase-query behavior when a user
# query happens to contain bare "AND", "OR", "NOT", "NEAR".
# FTS5 reserves these Boolean operator KEYWORDS at the token level.
_BOOLEAN_KW_RE = re.compile(r"(?<!\w)(AND|OR|NOT|NEAR)(?!\w)")
def _sanitize_query(text: str) -> str:
"""Reduce a natural-language query to an FTS5 OR-of-tokens query.
Two transformations:
1. Non-alphanumeric → space (drops punctuation; "10.9?" becomes
"10 9"). Lets us handle versions, parens, question marks, etc.
without inviting FTS5 parse errors.
2. Boolean keywords stripped (FTS5 reserves AND/OR/NOT/NEAR).
3. Tokens explicitly OR'd. FTS5's default is AND-of-tokens — for
any non-trivial natural-language query that means zero hits
(no chunk contains every word). OR semantics is what we want:
BM25 already weights documents containing more query terms
higher, so we don't lose precision, but we DO gain recall.
See ``crop-chem-docs`` for the rationale; same transformation
applies here. OR semantics maximizes recall — BM25 already
weights documents with more query-term matches higher.
"""
cleaned = _KEEP_RE.sub(" ", text)
cleaned = _BOOLEAN_KW_RE.sub(" ", cleaned)
@@ -113,9 +75,9 @@ def _where_to_sql(where: dict | None) -> tuple[str, list[Any]]:
Accepts the same shapes ``docs_mcp.server._build_where`` produces:
None → ("", [])
{"version": "10.9"} → ("AND m.version = ?", ["10.9"])
{"$and": [{...}, {...}]} → ("AND m.X = ? AND m.Y = ?", [...])
None → ("", [])
{"crop": "corn"} → ("AND m.crop = ?", ["corn"])
{"$and": [{...}, {...}]} → ("AND m.X = ? AND m.Y = ?", [...])
Unknown keys are silently dropped (defensive — better to over-match
than to crash on a filter we don't know).
@@ -158,35 +120,32 @@ class BM25Index:
def build(self, records: list[dict]) -> int:
"""Rebuild the index from scratch from `records`.
`records` is the same list ``rag.index.page_records`` produces:
`records` is the same list ``rag.index.variety_records`` produces:
``[{"id": ..., "text": ..., "metadata": {...}}, ...]``. Bulk
insert wrapped in a transaction — single-digit seconds for the
full 73k-chunk corpus.
insert wrapped in a transaction.
"""
self.db_path.parent.mkdir(parents=True, exist_ok=True)
# Drop and recreate. Idempotent rebuild.
if self.db_path.exists():
self.db_path.unlink()
with sqlite3.connect(self.db_path) as con:
con.executescript(self._schema_sql())
con.executemany(
"INSERT INTO chunks_meta (id, bundle_id, page_id, version, "
"platform, product, ordinal) VALUES (?, ?, ?, ?, ?, ?, ?)",
"INSERT INTO chunks_meta "
"(id, source, vendor, brand, crop, source_key, ordinal) "
"VALUES (?, ?, ?, ?, ?, ?, ?)",
[
(
r["id"],
r["metadata"].get("bundle_id") or "",
r["metadata"].get("page_id") or "",
r["metadata"].get("version") or "",
r["metadata"].get("platform") or "",
r["metadata"].get("product") or "",
r["metadata"].get("source") or "",
r["metadata"].get("vendor") or "",
r["metadata"].get("brand") or "",
r["metadata"].get("crop") or "",
r["metadata"].get("source_key") or "",
int(r["metadata"].get("ordinal") or 0),
)
for r in records
],
)
# Populate the FTS5 contentless-ish table by rowid. We populated
# chunks_meta first; rowids align with insertion order.
con.executemany(
"INSERT INTO chunks_fts (rowid, text) VALUES (?, ?)",
[
@@ -210,15 +169,12 @@ class BM25Index:
FTS5's bm25() returns NEGATIVE numbers — more relevant docs have
smaller (more negative) scores. We order ASC so the first row is
the most relevant. Callers that need a "rank" should enumerate
the returned list.
the most relevant.
"""
sanitized = _sanitize_query(text)
if not sanitized:
return []
where_sql, params = _where_to_sql(where)
# FTS5 MATCH wants the unaliased table name on its left, so we use
# chunks_fts (no alias) and JOIN by rowid against chunks_meta.
sql = (
"SELECT m.id, bm25(chunks_fts) AS score "
"FROM chunks_fts "
@@ -232,17 +188,13 @@ class BM25Index:
cur = con.execute(sql, [sanitized, *params, n])
return [(row[0], float(row[1])) for row in cur.fetchall()]
except sqlite3.OperationalError as e:
# FTS5 syntax error (rare after sanitization) or db missing.
# Caller decides whether to fall back to dense-only.
log.warning("bm25 query failed (%s); query=%r", e, sanitized[:80])
return []
def exists(self) -> bool:
"""Cheap probe — does the index file exist on disk?"""
return self.db_path.exists()
def count(self) -> int:
"""Number of chunks indexed. 0 if the db is missing or empty."""
if not self.exists():
return 0
try:
@@ -257,18 +209,19 @@ class BM25Index:
def _schema_sql() -> str:
return """
CREATE TABLE chunks_meta (
rowid INTEGER PRIMARY KEY AUTOINCREMENT,
id TEXT UNIQUE NOT NULL,
bundle_id TEXT,
page_id TEXT,
version TEXT,
platform TEXT,
product TEXT,
ordinal INTEGER
rowid INTEGER PRIMARY KEY AUTOINCREMENT,
id TEXT UNIQUE NOT NULL,
source TEXT,
vendor TEXT,
brand TEXT,
crop TEXT,
source_key TEXT,
ordinal INTEGER
);
CREATE INDEX idx_meta_version ON chunks_meta(version);
CREATE INDEX idx_meta_platform ON chunks_meta(platform);
CREATE INDEX idx_meta_bundle ON chunks_meta(bundle_id);
CREATE INDEX idx_meta_source ON chunks_meta(source);
CREATE INDEX idx_meta_crop ON chunks_meta(crop);
CREATE INDEX idx_meta_brand ON chunks_meta(brand);
CREATE INDEX idx_meta_source_key ON chunks_meta(source_key);
CREATE VIRTUAL TABLE chunks_fts USING fts5(
text,
+298 -100
View File
@@ -1,126 +1,324 @@
"""Markdown chunker — paragraph-aware, ~400-600 token target.
"""Chunker for seed-variety corpus.
Adjust the chunking strategy per product if your page format differs
significantly from prose. The output shape (id, text, metadata) is
fixed by the downstream Chroma + BM25 indexing in rag/index.py — don't
change that.
Each variety becomes ONE chunk by default. Variety pages are small
(typically 2-3 KB of useful signal) and nomic-embed-text handles up
to ~8 K tokens cleanly. Splitting a variety across chunks dilutes
the named-rating embeddings (e.g. "SCN resistance 7") that farmers
search by — keep them together.
The key knob you'll tune per product is chunk-0. Dense retrieval lands
on chunk 0 first for most queries. Make it a synthetic chunk built
from:
The chunk text is a synthetic preamble assembled deterministically
from the sidecar JSON. Every value in the chunk text comes verbatim
from the source. The framing words ("Disease ratings (1-9, 9=best):",
"Maturity group:", etc.) are template glue — *we add structure, we
do NOT add facts*. Given the same sidecar, this chunker always
produces the same chunk text. That's the anti-hallucination
contract: the retriever can never surface a rating value that
wasn't in the source.
- the page title (as natural-language H1)
- a 1-sentence task description (you'll have to generate this — for
pages that already have a "## Overview" or "## Introduction" the
first sentence usually works)
- a keyword bag of important terms (filenames, API names, error
codes — the rare technical tokens that BM25 lights up on)
Metadata is flattened to Chroma-safe primitives (str/int/float/bool):
Without a rich chunk 0, dense retrieval gets dominated by the much
larger prose body, and short pages (script examples, reference cards)
get buried.
source "bayer_seeds"
source_key "dekalb-dkc075-70rib"
vendor "Bayer"
brand "DEKALB"
crop "corn" | "soybeans" | "wheat"
product_name "DKC075-70RIB BRAND BLEND"
product_id canonical full id
source_url the variety's page URL
rm corn RM as int when parseable (else absent)
mg soy MG as float when parseable (else absent)
release_year int when known
trait_codes_csv comma-separated trait codes for substring search
rating_scale "1-9 (9 = best)" — chunker should ALWAYS attach
this so downstream code can detect a flip
ordinal chunk index within variety (0-based)
Lists like ``regional_recommendations`` and the full per-rating dicts
do NOT fit Chroma's metadata constraints — they stay in the sidecar
JSON, surfaced by ``get_page`` / ``lookup_variety``.
"""
from __future__ import annotations
import json
import re
from pathlib import Path
from typing import Iterator
# Approximate token estimate from char count. Tunable — set per
# embedder if the default 4 chars/token is wrong.
CHARS_PER_TOKEN = 4
TARGET_TOKENS = 500
TARGET_CHARS = TARGET_TOKENS * CHARS_PER_TOKEN
# Rating-group classification. The source publishes characteristics
# grouped by label; we map those labels to one of three buckets in
# the chunk preamble so retrieval gets coherent text. Group labels not
# listed here fall into "other" and are still emitted, just in their
# own section.
DISEASE_GROUP_LABELS = {
"DISEASE RATINGS",
"PEST AND DISEASE RESISTANCE",
}
AGRONOMIC_GROUP_LABELS = {
"GROWTH",
"HARVEST",
"PRODUCTION",
"KEY CHARACTERISTICS",
"QUALITY",
}
MANAGEMENT_GROUP_LABELS = {
"MANAGEMENT",
"HERBICIDE",
"SENSITIVITY",
"PLANT DESCRIPTION",
}
def estimate_tokens(text: str) -> int:
return max(1, len(text) // CHARS_PER_TOKEN)
def _parse_rm(value: object) -> int | None:
"""Best-effort RM-days int. Returns None if not a clean integer
(e.g. wheat's qualitative 'Early'/'Medium-Early' values)."""
if value is None:
return None
s = str(value).strip()
if not s:
return None
try:
# Handle floats stored as strings ("105.0") and the trailing
# tenths sometimes seen on early corn ("75").
return int(float(s))
except ValueError:
return None
def split_paragraphs(md: str) -> list[str]:
"""Split markdown into paragraph-ish blocks.
def _parse_mg(value: object) -> float | None:
"""Best-effort MG float. Soy MGs go from 00 to 9.0 with one decimal."""
if value is None:
return None
s = str(value).strip()
if not s:
return None
try:
return float(s)
except ValueError:
return None
Keeps fenced code blocks together (don't slice through ```).
Headings start new paragraphs.
def _format_items(items: list[dict]) -> str:
"""Render `[{characteristic, value}, ...]` to a compact inline list."""
out: list[str] = []
for it in items:
ch = (it.get("characteristic") or "").strip()
v = (it.get("value") or "").strip()
if ch and v:
out.append(f"{ch} {v}")
elif ch:
out.append(f"{ch}")
return ", ".join(out)
def _render_variety_chunk(sidecar: dict) -> str:
"""Build the dense preamble for one variety from its sidecar JSON.
Faithful to source: every numeric/categorical *value* is verbatim
from ``sidecar``. The only generated text is the framing language.
"""
blocks: list[str] = []
current: list[str] = []
in_fence = False
for line in md.splitlines(keepends=True):
stripped = line.strip()
if stripped.startswith("```"):
in_fence = not in_fence
current.append(line)
lines: list[str] = []
# ---- Identity line --------------------------------------------------
name = sidecar.get("product_name") or sidecar.get("source_key") or ""
brand = (sidecar.get("brand") or "").strip()
vendor = sidecar.get("vendor") or ""
crop = (sidecar.get("crop") or "").strip()
crop_label = crop.capitalize() if crop else ""
ident = f"# {name}"
sub = " ".join(filter(None, [
f"({brand.title()} {crop_label} variety, {vendor})" if brand and crop_label and vendor else "",
]))
lines.append(ident)
if sub:
lines.append("")
lines.append(sub)
# ---- Identity body --------------------------------------------------
facts: list[str] = []
rm = sidecar.get("relative_maturity")
mg = sidecar.get("maturity_group")
wc = sidecar.get("wheat_class")
if crop == "corn" and rm:
facts.append(f"Relative maturity {rm}")
elif crop == "soybeans" and mg:
facts.append(f"Maturity group {mg}")
elif crop == "wheat":
if rm:
facts.append(f"Maturity {rm}")
if wc:
facts.append(f"Wheat class {wc}")
traits = sidecar.get("trait_stack") or []
trait_descs = sidecar.get("trait_descriptions") or []
if traits:
if trait_descs:
facts.append(
"Trait stack: "
+ ", ".join(traits)
+ " ("
+ "; ".join(trait_descs)
+ ")"
)
else:
facts.append("Trait stack: " + ", ".join(traits))
if sidecar.get("release_year"):
facts.append(f"Released {sidecar['release_year']}")
if facts:
lines.append("")
lines.append(". ".join(facts) + ".")
# ---- Positioning ----------------------------------------------------
pos = (sidecar.get("positioning_statement") or "").strip()
if pos:
lines.append("")
lines.append(f"Positioning: {pos}")
# ---- Ratings, bucketed for retrieval --------------------------------
scale = sidecar.get("_scale_direction") or "(scale direction not declared)"
groups = sidecar.get("characteristics_groups") or []
disease: list[dict] = []
agronomic: list[dict] = []
management: list[dict] = []
other: list[tuple[str, list[dict]]] = []
for g in groups:
label = (g.get("label") or "").upper().strip()
items = g.get("items") or []
if not items:
continue
if in_fence:
current.append(line)
continue
if stripped.startswith("#"):
if current:
blocks.append("".join(current).strip())
current = []
current.append(line)
continue
if not stripped and current and not "".join(current).strip().endswith("\n\n"):
current.append(line)
blocks.append("".join(current).strip())
current = []
continue
current.append(line)
if current:
blocks.append("".join(current).strip())
return [b for b in blocks if b]
if label in DISEASE_GROUP_LABELS:
disease.extend(items)
elif label in AGRONOMIC_GROUP_LABELS:
agronomic.extend(items)
elif label in MANAGEMENT_GROUP_LABELS:
management.extend(items)
else:
other.append((g.get("label") or "Other", items))
if disease:
lines.append("")
lines.append(f"Disease ratings ({scale}): {_format_items(disease)}.")
if agronomic:
lines.append("")
lines.append(f"Agronomic ratings ({scale}): {_format_items(agronomic)}.")
if management:
lines.append("")
lines.append(f"Management notes: {_format_items(management)}.")
for label, items in other:
lines.append("")
lines.append(f"{label.title()}: {_format_items(items)}.")
# ---- Strengths narrative --------------------------------------------
strengths = sidecar.get("strengths") or []
if strengths:
lines.append("")
lines.append("Strengths and management notes:")
for s in strengths:
s = (s or "").strip()
if s:
lines.append(f"- {s}")
# ---- Regional listings (compact, not the agronomist emails) ---------
rec = sidecar.get("regional_recommendations") or []
if rec:
names = sorted({
(r.get("product_list_name") or "").strip()
for r in rec
if (r.get("product_list_name") or "").strip()
})
if names:
lines.append("")
lines.append("Listed in regional seed guides: " + "; ".join(names) + ".")
# ---- Provenance footer (must always be in the chunk text so it
# can never be lost between retrieval and LLM rendering) --------
urls = sidecar.get("source_urls") or []
if urls:
lines.append("")
lines.append(f"Source: {urls[0]}")
return "\n".join(lines).strip() + "\n"
def chunks_from_page(
text: str,
page_id: str,
metadata: dict,
def _flat_metadata(sidecar: dict) -> dict:
"""Distil sidecar into Chroma-safe metadata (primitives only)."""
md: dict = {
"source": sidecar.get("source") or "",
"source_key": sidecar.get("source_key") or "",
"vendor": sidecar.get("vendor") or "",
"brand": sidecar.get("brand") or "",
"crop": sidecar.get("crop") or "",
"product_name": sidecar.get("product_name") or "",
"product_id": sidecar.get("product_id") or "",
"source_url": (sidecar.get("source_urls") or [""])[0],
"rating_scale": sidecar.get("_scale_direction") or "",
}
rm = _parse_rm(sidecar.get("relative_maturity"))
mg = _parse_mg(sidecar.get("maturity_group"))
if rm is not None:
md["rm"] = rm
if mg is not None:
md["mg"] = mg
ry = sidecar.get("release_year")
if isinstance(ry, int):
md["release_year"] = ry
traits = sidecar.get("trait_stack") or []
if traits:
# Comma-delimited for partial-match / human eyeballing.
# Bracket-padded so `LIKE '%,XF,%'` finds whole tokens.
md["trait_codes_csv"] = "," + ",".join(traits) + ","
if sidecar.get("wheat_class"):
md["wheat_class"] = sidecar["wheat_class"]
return md
def chunks_from_variety(
sidecar_path: Path | str,
*,
md_path: Path | str | None = None,
) -> Iterator[dict]:
"""Yield chunk dicts ready for index.py to upsert.
"""Yield chunk dict(s) for one variety. Currently emits exactly one.
The synthetic chunk 0 is the per-product customization point. The
default below is a simple title + body-first-paragraph; rewrite
for richer retrieval signal (see module docstring).
Args:
sidecar_path: path to the variety's JSON sidecar.
md_path: ignored (the chunker rebuilds from sidecar); kept
in the signature in case a future split-chunker
wants the rendered body.
"""
paragraphs = split_paragraphs(text)
if not paragraphs:
return
# ----- Chunk 0: synthetic anchor for dense retrieval ---------
title = metadata.get("title") or page_id
first_para = next((p for p in paragraphs if not p.startswith("#")), "")
chunk0_body = (
f"# {title}\n\n"
f"{first_para[:300]}"
# TODO per product: append a keyword bag here (filenames,
# API names, error codes) for BM25 + dense joint coverage.
)
sidecar = json.loads(Path(sidecar_path).read_text(encoding="utf-8"))
text = _render_variety_chunk(sidecar)
meta = _flat_metadata(sidecar)
chunk_id = f"{meta['source']}::{meta['source_key']}::0"
yield {
"id": f"{metadata['bundle_id']}::{page_id}::0",
"text": chunk0_body,
"metadata": {**metadata, "ordinal": 0},
"id": chunk_id,
"text": text,
"metadata": {**meta, "ordinal": 0},
}
# ----- Body chunks: pack paragraphs up to TARGET_CHARS -------
ordinal = 1
buf: list[str] = []
buf_chars = 0
for p in paragraphs:
if buf_chars + len(p) > TARGET_CHARS and buf:
yield {
"id": f"{metadata['bundle_id']}::{page_id}::{ordinal}",
"text": "\n\n".join(buf),
"metadata": {**metadata, "ordinal": ordinal},
}
ordinal += 1
buf = []
buf_chars = 0
buf.append(p)
buf_chars += len(p)
if buf:
yield {
"id": f"{metadata['bundle_id']}::{page_id}::{ordinal}",
"text": "\n\n".join(buf),
"metadata": {**metadata, "ordinal": ordinal},
}
# ----- Backwards-compat shim for the template's index.py -------------------
#
# The template's ``rag.index.page_records`` calls
# ``chunks_from_page(md, page_id, base_meta)`` which doesn't know about
# sidecar JSON. We accept that signature but ignore it — index.py has
# been updated to use ``chunks_from_variety`` directly, and this shim
# is here only so a stray import of the old name doesn't break.
#
def chunks_from_page(text: str, page_id: str, metadata: dict) -> Iterator[dict]:
"""Deprecated for seed-mcp; prefer ``chunks_from_variety``."""
# Best-effort: if metadata carries a sidecar_path, dispatch.
sidecar_path = metadata.get("_sidecar_path")
if sidecar_path:
yield from chunks_from_variety(sidecar_path)
return
# Fallback — emit a single chunk of the raw markdown with whatever
# metadata we have. Better than crashing if someone calls this.
chunk_id = f"{metadata.get('source','unknown')}::{page_id}::0"
yield {
"id": chunk_id,
"text": text,
"metadata": {**metadata, "ordinal": 0},
}
+25 -38
View File
@@ -1,15 +1,19 @@
"""Build Chroma (and optionally BM25) indexes from corpus on disk.
"""Build Chroma (and BM25) indexes from the seed corpus on disk.
Reads `corpus/<bundle>/<page>.{md,json}`, chunks each page, upserts
into Chroma. With --rebuild, drops + recreates the collection (clean
state). With --bm25-only, skips Chroma and rebuilds only the FTS5
index — useful for fast iteration when chunking didn't change.
Reads ``corpus/<source>/<source_key>.json`` sidecars, chunks each
variety via ``rag.chunk.chunks_from_variety``, upserts into Chroma.
With ``--rebuild``, drops + recreates the collection (clean state).
With ``--bm25-only``, skips Chroma and rebuilds only the FTS5 index
— useful for fast iteration when the chunker didn't change.
Collection name is ``<PRODUCT_NAME>_docs`` (default: ``crop_seed_docs``).
Override via the PRODUCT_NAME env var.
"""
from __future__ import annotations
import argparse
import json
import logging
import os
import time
from pathlib import Path
from typing import Iterator
@@ -17,7 +21,7 @@ from typing import Iterator
import chromadb
from chromadb.config import Settings
from .chunk import chunks_from_page
from .chunk import chunks_from_variety
from .embeddings import embedding_function
log = logging.getLogger(__name__)
@@ -27,39 +31,21 @@ ROOT = Path(__file__).resolve().parent.parent
CORPUS = ROOT / "corpus"
CHROMA_DIR = ROOT / "chroma"
# Collection name — convention: <product>_docs. Override via env if needed.
import os
PRODUCT_NAME = os.environ.get("PRODUCT_NAME", "myproduct")
PRODUCT_NAME = os.environ.get("PRODUCT_NAME", "crop_seed")
COLLECTION = f"{PRODUCT_NAME}_docs"
def page_records() -> Iterator[dict]:
"""Walk corpus/, yield chunks for every page."""
def variety_records() -> Iterator[dict]:
"""Walk ``corpus/<source>/<source_key>.json``, yield one chunk per
variety."""
if not CORPUS.exists():
log.error("corpus/ doesn't exist; run the scraper first")
log.error("corpus/ doesn't exist; run a scraper first")
return
for bundle_dir in sorted(CORPUS.iterdir()):
if not bundle_dir.is_dir() or bundle_dir.name.startswith("."):
for source_dir in sorted(CORPUS.iterdir()):
if not source_dir.is_dir() or source_dir.name.startswith("."):
continue
for md_path in sorted(bundle_dir.glob("*.md")):
page_id = md_path.stem
sidecar = md_path.with_suffix(".json")
if not sidecar.exists():
log.warning("skipping %s — no JSON sidecar", md_path)
continue
md = md_path.read_text()
meta = json.loads(sidecar.read_text())
# Surface common filter fields at the chunk-metadata level
# so Chroma's `where` filter can use them.
base_meta = {
"bundle_id": bundle_dir.name,
"page_id": page_id,
"title": meta.get("title") or "",
"version": meta.get("version") or "",
"platform": meta.get("platform") or "",
"product": meta.get("product") or "",
}
yield from chunks_from_page(md, page_id, base_meta)
for sidecar_path in sorted(source_dir.glob("*.json")):
yield from chunks_from_variety(sidecar_path)
def upsert_to_chroma(records: list[dict]) -> int:
@@ -67,7 +53,7 @@ def upsert_to_chroma(records: list[dict]) -> int:
path=str(CHROMA_DIR),
settings=Settings(anonymized_telemetry=False),
)
# Drop + recreate for --rebuild semantics
# Drop + recreate for --rebuild semantics.
try:
client.delete_collection(COLLECTION)
except Exception:
@@ -101,8 +87,11 @@ def main() -> int:
log.info("reading corpus from %s", CORPUS)
t0 = time.time()
records = list(page_records())
records = list(variety_records())
log.info("loaded %d chunks in %.1fs", len(records), time.time() - t0)
if not records:
log.error("no chunks — is corpus/ populated?")
return 1
if args.bm25_only:
from .bm25 import BM25Index
@@ -118,8 +107,6 @@ def main() -> int:
n = upsert_to_chroma(records)
log.info("chroma: %d chunks in %.1fs", n, time.time() - t_c)
# Build BM25 too — see PLAN.md Phase 8. Safe to remove this block
# for products that don't need hybrid retrieval.
try:
from .bm25 import BM25Index
t_b = time.time()