Perq

Match scraped documents against large sets of saved search queries in a single Python process. Build the query index once, then match each extracted product, listing, job, or article as it arrives. Matching runs as a local function call inside your scraper, with no search service to deploy or network request per document.

Perq, formerly PSearch, is an embedded percolator for prospective search, reverse search, saved-search alerts, and document routing. It uses rare-clause candidate selection and bitmasks so that each document only needs to complete the checks for candidate queries. A large saved-query set can be served by one process. The core package has no runtime dependencies and supports Python 3.11–3.14.

In the checked synthetic benchmarks, one process matched documents against 100,000 saved queries at 5,801 documents/second, and against one million at 599 documents/second, using MemoryStore. The million-query run used 842 MiB peak process memory. These figures cover short, pre-tokenized documents drawn from a uniform vocabulary; longer and common-term workloads are reported too.

Install

Install from PyPI:

python -m pip install perq
# Optional persistent key-value backend:
python -m pip install 'perq[lmdb]'

The distribution and import names are perq. Wheels and source archives are also available from GitHub releases.

Match a scraped item

from perq import Document, MemoryStore, Query, QueryMatcher, RangeFilter, index

# Build once when the scraper starts; reuse this matcher for every item.
with MemoryStore() as store:
    index(
        [
            Query(
                "camera-alert",
                [("sony",), ("a7", "a7iii")],
                filters=[RangeFilter("price_eur", lte=1200)],
            ),
            Query.from_text("lens-alert", "canon lens"),
        ],
        store,
    )
    matcher = QueryMatcher(store)

    # Feed fields already extracted by your scraper.
    document = Document.from_text(
        "Sony A7 III mirrorless camera, used body", values={"price_eur": [950]}
    )
    assert list(matcher.matches(document)) == ["camera-alert"]

The first query means sony AND (a7 OR a7iii) AND price_eur ≤ 1200. IDs identify saved searches or subscriptions and can be strings or integers. Results follow query input order. See the streaming scraper example for matching an iterator of extracted items. Perq handles the matching step; your scraper handles fetching, extraction, and delivery of alerts.

Use SQLiteStore("queries.sqlite") for portable persistence, or optional LMDBStore("queries.lmdb") for read-heavy local indexes. Each index(...) call atomically replaces the whole query set. A failed build preserves the old index; index([], store) clears it. This release does not expose incremental updates.

Match JSONL from a terminal or agent

perq build --index queries.sqlite --queries examples/queries.jsonl
perq match --index queries.sqlite --documents examples/documents.jsonl
perq inspect --index queries.sqlite

match emits one JSON object per document:

{"id": "doc-1", "matches": ["data-alert"]}
{"id": "doc-2", "matches": ["python-route", "price-only"]}
{"id": "doc-3", "matches": []}

Input defaults to stdin. Diagnostics go to stderr. See the checked query and document schemas and the agent usage guide.

Performance and ease of use

Perq is a good fit for scraping workloads that reuse many saved searches as new items arrive: product and price monitoring, classifieds, job alerts, and article filtering. Query count alone does not require a separate matching service. Start with MemoryStore when the index fits in RAM, and reuse it for the lifetime of the scraping process. Index construction is paid once per query-set replacement.

The Elasticsearch setup and batching behavior are described in its percolator documentation. Perq's direct advantage for a Python scraper is its small deployment and integration cost. The timings above compare the same three documents against a straightforward Python matcher; the benchmark report gives the workload, sample sizes, and cases where the advantage is smaller.

Query support

Matching is exact after normalization. There is no semantic similarity, ranking, stemming, phrase/proximity operator, negation, or general query-string parser. Queries support up to 64 AND clauses, each containing one or more OR terms. Numeric-only queries are supported and evaluated for every document. Match-all queries and empty OR clauses are rejected.

Storage and performance

Disk builds stage queries and sort postings in a temporary SQLite file. They retain term frequencies and one term's posting list in memory, rather than all posting tuples. MemoryStore additionally retains the finished index and, during replacement, the old generation. Staging requires temporary disk space. This is not a guarantee of bounded total memory for every workload.

Performance depends on document length, term distribution, Boolean breadth, and the number of emitted matches. See benchmark methodology and results and the 2026 backend decision. Broad queries and common terms can still require work proportional to the query count.

Development

python -m pip install '.[dev,lmdb]'
python -m pytest
ruff check .
ruff format --check .
python -m build
twine check dist/*

Documentation examples, CLI behavior, JSON schemas, failure recovery, and randomized reference comparisons are tested. See usage, migration notes, contributing, and the changelog. Licensed under MIT; contributors are listed in AUTHORS.