Search modelica docs

Search Modelica & System Modeler Documentation (BM25)

A self-contained, offline documentation search. It ranks passages with BM25 (pure Python standard library — no third-party packages, no network, no API) over three prebuilt corpora and returns the matching chunks with their citation URLs. Use it to ground answers in real documentation instead of recalling from memory.

The three corpora (and how to route)

The content covers the Modelica language, the System Modeler tool, and the Modelica Standard Library — pick the corpus by what the user means:

--corpus value	Covers	Backing corpora
modelica	The Modelica language — syntax, semantics, the normative specification	spec
systemmodeler (or sm)	Wolfram System Modeler, the tool — building/simulating models, GUI, tutorials, what's-new	docs
msl (or library)	The Modelica Standard Library — component/block docs, exact parameter names, example models, annotation-stripped source	msl
all (default)	Everything	spec + docs + msl
spec	Modelica Language Specification (normative)	spec
docs	System Modeler documentation	docs

Routing guidance:

• Language questions ("how does when work", "is this valid Modelica", "what's the rule for...") → --corpus modelica.
• Tool questions ("how do I simulate / plot / use the GUI in System Modeler", "what's new in 15") → --corpus systemmodeler.
• Library questions ("which block gives me an anti-windup PID", "what are the parameters of Inertia", "show an example wiring a spring-damper") → --corpus msl. Also use this while writing a model with MSL components to confirm parameter names and connect() patterns.
• When unsure, use all (the default) — a question often spans corpora, and the results are tagged so you can see where each came from. But msl is by far the largest corpus (~4,700 chunks vs ~900 for the other two), so if you know the question is about the language or the tool, route to that corpus — it keeps library chunks from crowding out the answer.

Requirements

Python 3 only. No pip install needed — the index uses just the standard library and the prebuilt JSON data ships inside this skill (docsearch/data/). Run commands from this skill's own directory (the folder containing this SKILL.md) so the docsearch package imports. Use python instead of python3 on Windows if that's what's on PATH.

Commands

cd "<skill-dir>"     # the folder containing this SKILL.md

# Search (default k=5). Route with --corpus.
python -m docsearch.main --query "how do I write a when statement" --corpus modelica
python -m docsearch.main --query "simulate a fluid tank" --corpus systemmodeler --k 8
python -m docsearch.main --query "anti-windup PID controller output limits" --corpus msl

# Read the FULL text of a specific result (paste its exact url).
python -m docsearch.main --show "https://specification.modelica.org/maint/3.6/statements-and-algorithm-sections.html#4gfvu7z5m8rutn1wqr3gsd7mr"
python -m docsearch.main --show "modelica://Modelica.Blocks/LimPID"

# Machine-readable output for programmatic use.
python -m docsearch.main --query "array construction" --json

# List corpora + aliases / counts.
python -m docsearch.main --corpora

# (Maintainer, repo checkout only) measure retrieval quality against the gold sets.
python -m docsearch.main --eval --k 5

Flags: --corpus (route, default all), --k (results, default 5), --full (print whole chunks instead of 600-char previews), --json, --show <url>.

Workflow

1. Route. Decide modelica vs systemmodeler vs all from the question (see the table). Default to all if it could be either.
2. Search. Run --query with the user's question. BM25 is lexical, so if the first try is weak, reformulate with domain keywords (e.g. add when clause discrete event, connector flow potential, algorithm section) and search again, and/or raise --k. A second targeted query usually finds it.
3. Read. The preview is the first 600 chars. To quote or reason carefully, pull the full passage with --show <url> (or re-run with --full).
4. Answer and cite. Base the answer on the retrieved text and cite the source URL(s) so the user can verify. Each result is tagged with its corpus; follow the citation guidance the tool prints at the top of search results:
   • Specification (spec) text is authoritative/normative. Text in brackets [...] is non-normative (explanatory only) — don't present it as a rule. When the answer quotes or is derived from spec excerpts, end it with this attribution: Derived from the Modelica® Language Specification. Copyright © Modelica Association. Used under CC BY-SA 4.0.
   • System Modeler docs (docs) are tutorials / user guide / release notes, © Wolfram Research — cite the reference.wolfram.com source URL.
   • MSL reference (msl) entries are class docs plus annotation-stripped MSL 4.0 source. The code is authoritative for parameter names, defaults, and connect() wiring. Entries are keyed by leaf class name as modelica://<TopPackage>/<Name> pseudo-URLs — cite the class name (e.g. Modelica.Blocks.Continuous.LimPID if the doc text gives the full path), not the pseudo-URL, since it is not a resolvable link.
5. If nothing relevant comes back after reformulating, say so rather than inventing an answer — that is the whole point of this skill.

Notes

• Retrieval is deterministic and local; identical queries return identical results. There is no learning or external call.
• The --eval gold sets (docsearch/data/eval/*.jsonl) are query→URL pairs used only to measure recall@k / MRR when tuning; normal use never reads them, and they are not shipped in the release bundle (repo checkout only).