harvester-query

Specter's Harvester is a FastAPI service that fronts two Chroma collections:

| Collection | Contents | When the API uses it | |---|---|---| | specter_statutes | 384 US motor-vehicle statutes (CA, TX, FL, PA, IL, OH) | default — any US-jurisdiction or generic query | | canada_pi_cases | 340 real Canadian PI cases (CanLII, with damages, injury type, plaintiff_won, deciding_factor) | auto-routed when the query mentions Canada / Ontario / Quebec / Toronto / ONSC / ONCA / canlii / etc., or when ?jurisdiction=Canada or ?state=CA-CAN is passed |

Every citation the user sees should come from the Harvester (or, on miss, from a verified web source). Never fabricate.

When to use this skill

• Always for fact patterns mentioning Canada or a Canadian province / city. The Canada collection is indexed on facts of decided cases, so messy fact patterns ("child got hurt outside Target in Ontario", "slip and fall on icy sidewalk in Toronto") are EXACTLY what it's good at. Run search first; ask follow-ups only after showing what hit.
• User asks for a specific citation: "what's Cal. Veh. Code § 23152(a)?" → lookup_citation
• User asks a semantic / topical question: "DUI statutes in California", "laws about texting while driving" → search
• User asks an open question and you want top-k candidates: "any state with a hit-and-run analog to CVC 20001?" → ask
• User uses a synonym ("speeding", "drunk driving", "texting") — map it to one of the 17 canonical contributing-factor categories before passing as a factor filter (see categories below).

When NOT to use it

• Pure procedural / case-strategy questions ("how do I file?") — the Harvester only holds statutes (US) and case law (Canada).
• Greetings, identity questions, persona check-ins — answer in voice without calling the API.
• Questions that don't require a citation (Specter can answer general PI-law-shape questions briefly without one, but must not invent a citation to back them up).

Important: don't gate behind clarification

For Canadian fact-pattern queries especially, call search first with whatever the user said, even if it's vague. Show the top hits, THEN ask any follow-ups. The Canada corpus is fact-pattern-indexed — thin queries still return relevant cases. Demanding more facts before searching wastes the corpus.

How to call

The helper module sits next to this SKILL.md:

python
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent))  # if running ad hoc
from client import lookup_citation, search, ask, healthz, list_factors

In OpenClaw, prefer running it via exec:

bash
cd ~/Specter && python3 -m openclaw.skills.harvester_query.client lookup "Cal. Veh. Code § 23152(a)"
cd ~/Specter && python3 -m openclaw.skills.harvester_query.client search "drunk driving" --state CA --factor "DUI/DWI"
cd ~/Specter && python3 -m openclaw.skills.harvester_query.client ask "what statutes cover hit and run in CA"

Or from Python:

python
from openclaw.skills.harvester_query.client import lookup_citation, search, ask

lookup_citation(citation: str) -> dict | None

Exact match by citation string. Returns the full record or None if not in the Harvester.

search(q: str, *, state=None, factor=None, pi_only=False, k=10) -> list[dict]

Semantic search. state is a 2-letter code ("CA", "TX"). factor must be one of the 17 canonical categories.

ask(question: str, *, state=None, factor=None, k=8) -> list[dict]

Same as search but framed as a question. Use this when the user hands you a natural-language query.

healthz() -> dict

Cheap liveness check. Use to confirm the API is running before reporting a miss.

list_factors() -> list[str]

The 17 categories. Useful for synonym mapping in voice.

How to format results for WhatsApp

Specter's voice rules (see SOUL.md): 2–3 lines max by default. Cite, don't quote.

Good: > Cal. Veh. Code § 22350. Basic speed law — drive no faster than reasonable for conditions. > leginfo.legislature.ca.gov

Bad: > Of course! Pursuant to California Vehicle Code Section 22350, "no person shall drive a vehicle…" (wall of text)

Always include the source_url from the record. Never include a citation without one — per hackathon ground rules, records without a real source URL don't count.

Miss handling

If lookup_citation returns None or search/ask returns []:

1. Check healthz() — if the API is down, say so plainly.
2. If the API is up but the record isn't there, say: "Not in the Harvester yet." Then either:

• Offer to web-fetch from an authoritative source (legislatures, courts, bar associations only).
• Suggest the team ingest that jurisdiction.

Never paper over a miss with a guess.

The 17 contributing-factor categories (canonical, do not invent)

DUI/DWI
Driving Too Fast For Conditions
Failure to Maintain Lane
Failure to Obey Traffic Control Device
Failure to Use/Activate Horn
Failure to Yield at a Yield Sign
Failure to Yield the Right-of-Way
Fleeing a Police Officer
Fleeing the Scene of a Collision
Following Too Closely
Improper Lane of Travel
Improper Passing
Improper Starting
Improper Stopping
Improper Turning
Reckless Driving
Using a Wireless Telephone/Texting While Driving

If the user uses a synonym, map internally and answer in their words.