feat(decisions): support decisions modules and the decide query argument

[jfrancoa]

What

Client support for the decisions-* module type and the decide query argument, starting with decisions-typesafeai.

Server side: weaviate/weaviate#13477 (module type), weaviate/weaviate#13408 (decisions-typesafeai), weaviate/weaviate#13484 (decide property, proto and decisions-dummy). This PR carries the regenerated decisions_pb2 stubs from that proto. Support for decisions-openai comes in a stacked PR (#2189), and rerank=Decide(...) in #2198.

Collection configuration

from weaviate.classes.config import Configure, Reconfigure, DecisionsOrder

client.collections.create(
    "Tickets",
    reranker_config=Configure.Reranker.cohere(),  # optional: a reranker next to the decisions module
    decisions_config=Configure.Decisions.typesafeai(
        min_probability=0.5,               # rerank: drop results below this probability
        order=DecisionsOrder.PROBABILITY,  # rerank order: by probability instead of search order
    ),
)
collection.config.update(decisions_config=Reconfigure.Decisions.typesafeai(min_probability=0.7))
collection.config.get().decisions_config  # DecisionsConfig(decisions="decisions-typesafeai", model={...})

A collection has at most one reranker and at most one decisions module, and an update replaces the module of its own kind only. Configure.Decisions.custom(name, settings) covers a module the client does not know. The settings are model, base_url, max_documents, min_probability and order; the module batches requests by itself.

Queries

from weaviate.classes.query import Decision

res = collection.query.hybrid(
    "refund not received",
    decide=[
        Decision.predicate("angry", prop="text", instructions="Is the customer angry?"),
        Decision.choice("team", prop="text", instructions="Which team should handle this?",
                        options={"billing": "money and invoices", "support": None}),
        Decision.score("severity", prop="text", instructions="How severe is the complaint?",
                       levels=["low", "medium", "high"]),
    ],
)
for o in res.objects:
    o.decisions["angry"].probability      # 0.91
    o.decisions["team"].choice            # "billing", plus .probabilities and .confidence
    o.decisions["severity"].score         # 2.0 (position between the levels), plus .probabilities
    o.decisions["angry"].refused          # True when the model declined to answer

decide= is accepted by every search method and by fetch_objects, and can be combined with rerank=. It never drops or reorders an object. It is not available with group_by.

Tests

• Unit: configuration serialisation, a reranker next to a decisions module, update replacement within a kind, read-back, argument validation.
• Mock: a Search reply with every answer kind is parsed into DecisionAnswer; the request carries the questions.
• Integration (integration/test_collection_decisions.py): runs against the server's decisions-dummy module, whose answers are a function of the text; skips when the server lacks the module. Verified locally against a build of feat(modules): add the decide additional property weaviate#13484. ci/docker-compose.yml enables the module, so the tests start running once a server image ships it.

[orca-security-eu]

Orca Security Scan Summary

Status	Check	Issues by priority
Passed	Infrastructure as Code	0   0   0   0	View in Orca
Passed	SAST	0   0   0   0	View in Orca
Passed	Secrets	0   0   0   0	View in Orca
Passed	Vulnerabilities	0   0   0   0	View in Orca

[codecov-commenter]

Codecov Report

❌ Patch coverage is 83.08271% with 45 lines in your changes missing coverage. Please review.
✅ Project coverage is 88.95%. Comparing base (95b5d76) to head (f1a6600).
⚠️ Report is 219 commits behind head on main.

Files with missing lines	Patch %	Lines
integration/test_collection_decisions.py	30.15%	44 Missing ⚠️
weaviate/collections/grpc/query.py	95.23%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2188      +/-   ##
==========================================
+ Coverage   86.64%   88.95%   +2.31%     
==========================================
  Files         300      310      +10     
  Lines       23172    24692    +1520     
==========================================
+ Hits        20077    21965    +1888     
+ Misses       3095     2727     -368     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.