Summa Python client

Async Python client for the Summa gRPC search server.

Installation

pip install summa-client-python

Python 3.10 or newer is required.

Quick start

import asyncio

from summa_client_python import SummaClient


async def main():
    async with SummaClient("localhost:50051") as client:
        await client.create_index(
            "articles",
            """
            index articles {
                field id: text<raw> [primary, stored]
                field title: text<simple> [indexed, stored]
                field body: text<simple> [indexed, stored]
            }
            """,
        )

        indexed, error_count, errors = await client.index_documents(
            "articles",
            [
                {"id": "1", "title": "Hello World", "body": "First article"},
                {"id": "2", "title": "Summa Search", "body": "Fast retrieval"},
            ],
        )
        if error_count:
            raise RuntimeError(errors)
        print(f"Indexed {indexed} documents")

        await client.commit("articles")

        results = await client.search(
            "articles",
            query={"match": {"field": "title", "text": "hello"}},
            fields_to_load=["title", "body"],
        )
        for hit in results.hits:
            print(hit.address, hit.score, hit.fields)

        if results.hits:
            document = await client.get_document("articles", results.hits[0].address)
            print(document.fields if document else "document not found")


asyncio.run(main())

The context manager connects and closes the channel. Manual callers use await client.connect() / await client.close().

Index management

await client.list_indexes()
info = await client.get_index_info("articles")
await client.reorder("articles")
await client.retrain_vector_index("articles")

index_documents returns (indexed_count, error_count, errors); index_documents_stream takes an async iterable and returns counts. Inspect errors, then commit() to publish accepted work. Repeated values are lists; flat numeric lists are dense vectors, (dimension, weight) pairs are sparse vectors. See client types.

Delete and upsert documents

With a text primary key, delete by exact key or supply a complete replacement:

await client.delete_document("articles", "2")
await client.upsert_document("articles", {"id": "1", "title": "Updated title"})
await client.commit("articles")

delete_documents / upsert_documents return DocumentMutationResult with accepted_count and errors: [{index, error}]. Single-item helpers raise on rejection. Missing deletes succeed; upserts insert missing keys. Staged rows can be replaced/deleted again before commit; the latest accepted version wins. Optional content hashes skip unchanged writes.

Limits: 100,000 deletion keys / 8 MiB key bytes; 1,000 replacements / 32 MiB encoded protobuf, or one replacement / 200 MiB including the request envelope. Broker commits are atomic per partition. See mutation semantics.

Compact deleted rows

await client.force_merge("articles")  # Retain tombstones.
await client.force_merge("articles", compact=True)  # Physically remove deleted rows.

Compaction handles singleton segments and may change addresses and BM25 scores. Index info exposes num_docs, physical_num_docs, num_deleted_docs, and deleted_ratio. Use primary keys for durable identity.

Searching

search(index_name, query=..., limit=10, fields_to_load=[...]) accepts one query variant: term, match, phrase, boolean, sparse_vector, dense_vector, binary_dense_vector, boost, range, prefix, all, or fusion. See query types and the wire contract for options.

results = await client.search(
    "articles",
    query={
        "boolean": {
            "must": [{"match": {"field": "title", "text": "search"}}],
            "must_not": [{"term": {"field": "title", "term": "draft"}}],
        }
    },
    fields_to_load=["title"],
)

get_document(index_name, hit.address) uses the full segment/document address and returns None on NOT_FOUND.

Ranking diagnostics and recall traces

Search options include_rrf_scores=True and tracing=True default to false. RRF diagnostics describe organic branch nominations; traces retain bounded candidates and query trees, including hits outside the final page. Neither changes retrieval depth or ranking. Oversized exports fail explicitly.

For named branches, use l1={"formula": "0.2 * title + 0.8 * body + 3 * rrf"}. The formula is the only L1 scoring interface; old coefficient fields are removed. See candidate scoring for backfill, passage selection, expression limits, capability versions, and distributed behavior.

Deadlines and errors

RPCs accept timeout in seconds, overriding the constructor's default_timeout. gRPC failures raise grpc.aio.AioRpcError, except document NOT_FOUND as above. An expired mutation may already be staged, and an accepted commit continues after disconnection. Resolve the outcome before retrying replacements.

Development

From this directory:

uv sync --group dev --group test
uv run ruff check .
uv run ruff format --check .
uv run pytest tests/test_client_unit.py
uv run --group dev python generate_proto.py

Integration tests require target/debug/summa-server. Regenerate bindings after protocol changes.