Azure Search Documents

azure-search-documents — Post-Regeneration Customization Guide

After running tsp-client update, the generated code is a raw skeleton — not a shippable SDK. This skill tells you exactly which _patch.py files customize it and how to verify those customizations still work after regeneration.

The generator never touches _patch.py files, so your customizations survive regeneration. Only edit _patch.py files for SDK customizations; the sole exception is adding inline # pylint: disable= comments to generated files for pylint issues. _patch.py files import from generated modules that do change, so a renamed model, a new parameter, or a removed enum value can silently break them. The verification steps below catch those breaks.

Additional references

Use these sidecars for guidance that is required for release-ready changes but does not belong in the generated post-regeneration checklist.

• Release tracks, branch-dependent API surface, and CHANGELOG rules: references/release.md
• Test commands, live recording workflow, and test authoring rules: references/testing.md
• Sample authoring workflow, .http contract parity, and live-run setup: references/samples.md

Environment setup

Use the alias as a prefix for any command that needs the package environment, for example venv python, venv pip, venv azpysdk, or venv python -m pytest. Plain git, grep, and file commands do not need it.

Set-Alias venv .\.github\skills\azure-search-documents\scripts\Invoke-VenvCommand.ps1

Step 1: Run Regeneration

cd sdk/search/azure-search-documents

# Update tsp-location.yaml with the new spec commit SHA, then:
tsp-client update
# or: azsdk_package_generate_code

# Reapply temporary Python emitter workarounds, then verify they are present:
python .github/skills/azure-search-documents/scripts/apply_generator_workarounds.py
python .github/skills/azure-search-documents/scripts/apply_generator_workarounds.py --check

# If the API version changed, _metadata.json updates automatically;
# reconcile the hand-maintained ApiVersion enum in Step 3.

Step 2: Verify Imports in All _patch.py Files

Every non-empty _patch.py imports from generated modules. After regeneration, check those imports still resolve:

venv python -c "from azure.search.documents import SearchClient, SearchIndexingBufferedSender, ApiVersion, DEFAULT_VERSION, IndexDocumentsBatch"
venv python -c "from azure.search.documents.aio import SearchClient, SearchIndexingBufferedSender"
venv python -c "from azure.search.documents.indexes import SearchIndexClient, SearchIndexerClient"
venv python -c "from azure.search.documents.indexes.aio import SearchIndexClient, SearchIndexerClient"
venv python -c "from azure.search.documents.indexes.models import SearchField, SearchFieldDataType, SimpleField, SearchableField, ComplexField, KnowledgeBase"
venv python -c "from azure.search.documents.knowledgebases import KnowledgeBaseRetrievalClient"
venv python -c "from azure.search.documents.knowledgebases.aio import KnowledgeBaseRetrievalClient"

If any fail, a generated class/enum was renamed or removed. See references/customizations.md for the exact import each _patch.py depends on.

The _patch.py files that have customizations (others are empty boilerplate):

azure/search/documents/
├── _patch.py                          # SearchClient subclass, SearchIndexingBufferedSender, ApiVersion enum, DEFAULT_VERSION
├── _operations/_patch.py              # SearchItemPaged, search(), custom index_documents/413 splitting, continuation-token helpers
├── models/_patch.py                   # IndexDocumentsBatch, RequestEntityTooLargeError
├── aio/_patch.py                      # Async SearchClient, async SearchIndexingBufferedSender
├── aio/_operations/_patch.py          # AsyncSearchItemPaged, async search(), reuses sync helpers
├── indexes/_patch.py                  # SearchIndexClient, SearchIndexerClient (audience kwarg plumbing)
├── indexes/_operations/_patch.py      # Polymorphic delete/create-or-update, list_index_names, _convert_index_response
├── indexes/models/_patch.py           # SearchField(hidden), field builders, SearchIndexerDataSourceConnection overloads, KnowledgeBase, SearchFieldDataType aliases
├── indexes/aio/_patch.py              # Async SearchIndexClient/SearchIndexerClient mirrors
├── indexes/aio/_operations/_patch.py  # Async mirror of indexes operations mixin
├── knowledgebases/_patch.py           # KnowledgeBaseRetrievalClient (audience kwarg plumbing)
└── knowledgebases/aio/_patch.py       # Async KnowledgeBaseRetrievalClient

Step 3: Check ApiVersion

The generator targets the version in _metadata.json. The hand-maintained ApiVersion enum in azure/search/documents/_patch.py must include it.

# 1. See what API version the generator used
venv python -c "import json; print(json.load(open('_metadata.json'))['apiVersion'])"

# 2. See what versions the SDK currently advertises
grep -A 20 'class ApiVersion' azure/search/documents/_patch.py

# 3. Check the current default
grep 'DEFAULT_VERSION' azure/search/documents/_patch.py

If the generated API version is not in the enum:

1. Add the new member to ApiVersion in azure/search/documents/_patch.py (e.g., V2026_05_01_PREVIEW = "2026-05-01-preview").
2. Update DEFAULT_VERSION to point to the new member.
3. Update docstrings on the public client subclasses (SearchClient, SearchIndexClient, SearchIndexerClient, KnowledgeBaseRetrievalClient) that name the default version.

Verify the default round-trips:

venv python -c "from azure.search.documents._patch import ApiVersion, DEFAULT_VERSION; print(DEFAULT_VERSION.value)"

Step 4: Expose New Generated Methods Through _patch.py

New generated operations do NOT automatically surface in the user-expected shape. Diff what regeneration added:

git diff --name-only | grep "_operations\.py" | grep -v _patch
git diff azure/search/documents/indexes/_operations/_operations.py | grep -E '^\+\s+(async\s+)?def '
git diff azure/search/documents/_operations/_operations.py | grep -E '^\+\s+(async\s+)?def '

For each new method, pick the exposure path:

1. Pass-through via generated mixin inheritance — if the generated signature is already user-friendly, the _<Client>OperationsMixin subclass in _patch.py inherits it automatically. Verify the method name does NOT collide with an existing _patch.py override.

2. Override in the operations mixin — needed when the method requires custom paging (SearchItemPaged), request-body rewriting (_build_search_request), response conversion (_convert_search_result, _convert_index_response), or 413 batch splitting.

3. Polymorphic str-or-model wrapper — for delete_* on a named resource (index, synonym map, alias, indexer, data source, skillset, knowledge base, knowledge source). Accept either the str name or the resource model; forward e_tag + match_condition when a model is passed.

4. Create-or-update wrapper — for create_or_update_*. Forward prefer="return=representation", match_condition, etag, plus package-specific flags:

   • allow_index_downtime for create_or_update_index
   • skip_indexer_reset_requirement_for_cache for indexers
   • skip_indexer_reset / disable_cache_reprocessing_change_detection for skillsets

5. List projection wrapper — for list_*. Add a select parameter, a name-only projection via cls callback (list_index_names, list_indexer_names, list_skillset_names), or convert a generated projection type back to the canonical model via _convert_index_response.

6. Friendly-input wrapper — when a generated parameter takes a structured string the caller would otherwise assemble by hand, add an optional higher-level kwarg that builds it from simpler inputs. Keep the raw parameter accepted too.

7. Re-export via __all__ — any NEW symbol you add or override in _patch.py MUST be appended to that file's __all__. Otherwise patch_sdk() will not surface it.

   venv python -c "import azure.search.documents as m; print(sorted(m.__all__))"
   venv python -c "import azure.search.documents.indexes as m; print(sorted(m.__all__))"
   

Every sync wrapper needs a matching async mirror under aio/. Shared helpers (_build_search_request, _convert_search_result, _convert_index_response, _pack_continuation_token, _unpack_continuation_token) live in the sync _operations/_patch.py and are imported by async — do not duplicate them.

Step 5: Check for Model/Enum Changes That Affect Customizations

git diff azure/search/documents/models/_models.py azure/search/documents/models/_enums.py
git diff azure/search/documents/indexes/models/_models.py azure/search/documents/indexes/models/_enums.py

Watch for:

• Renamed enum values — indexes/models/_patch.py monkey-patches camelCase aliases onto SearchFieldDataType (String, Int32, Int64, Single, Double, Boolean, DateTimeOffset, GeographyPoint, ComplexType). All right-hand-side UPPER_CASE members must still exist.
• Changed model constructors — IndexDocumentsBatch, SearchField, SearchIndexerDataSourceConnection, KnowledgeBase are subclassed in _patch.py; base-class constructor changes break them.
• New fields on SearchResult — _convert_search_result() extracts @search.* metadata (score, reranker_score, highlights, captions, document_debug_info, reranker_boosted_score). New metadata fields must be added there.
• Changed SearchRequest — _build_search_request() constructs this model directly; new query parameters must be wired through, including any new pipe-delimited semantic encoding.
• SearchIndexResponse.semantic — used by _convert_index_response(); if the field is renamed, update the mapping.
• SearchField.retrievable — the hidden property in SearchField subclass is its inverse; if renamed, update the property.

Step 6: Run package validation

Run the full validation suite (mypy, pylint, and the rest) in one shot:

azsdk_package_run_check with checkType="All"

mypy.ini and the repo-level pylintrc already exclude generated internals, _vendor/, tests/, and samples/ — errors will originate in _patch.py customizations.

Step 7: Update Documentation and Samples

CHANGELOG

Follow the CHANGELOG rules in references/release.md. Add entries under:

• ### Features Added — new operations, models, parameters, API version support
• ### Breaking Changes — renamed/removed models, changed signatures, dropped API versions
• ### Bugs Fixed — fixes to _patch.py logic, pagination, encoding, batching

README

• Add usage examples for new client classes (KnowledgeBaseRetrievalClient etc.) or operations.
• Update "Key concepts" for new resource types.
• Update the listed API version in "Getting started" if DEFAULT_VERSION changed.

Customization Patterns Reference

Each pattern below exists in the codebase today. After regeneration, verify each still works. See references/customizations.md for the exhaustive file-by-file inventory.

Client Constructor Customizations (audience / parameter reorder)

• SearchClient in _patch.py (sync + async) reorders params to (endpoint, index_name, credential). The generated base uses (endpoint, credential, index_name).
• SearchIndexClient, SearchIndexerClient (indexes/_patch.py + async) and KnowledgeBaseRetrievalClient (knowledgebases/_patch.py + async) all pop the audience kwarg and convert it into credential_scopes=[audience + "/.default"].

If regeneration changes the base constructor signature (param order, required kwargs), update each subclass.

Custom Search Pagination

search() does not use standard Azure SDK paging. It uses SearchItemPaged / AsyncSearchItemPaged with a custom SearchPageIterator because:

• Search paginates via POST with nextPageParameters body, not GET with nextLink.
• First-page metadata (facets, count, coverage, answers, debug info) must be available before iteration starts.
• Results are converted to dicts with @search.* keys.

Continuation token is Base64-encoded JSON: {"apiVersion": ..., "nextLink": ..., "nextPageParameters": {...}}.

Shared helpers (_build_search_request, _convert_search_result, _pack_continuation_token, _unpack_continuation_token) live in sync _operations/_patch.py and are imported by async.

Pipe-Delimited Semantic Encoding

_build_search_request() encodes semantic parameters into pipe-delimited wire format:

query_answer="extractive", count=3         →  "extractive|count-3"
query_caption="extractive", highlight=True →  "extractive|highlight-true"

New semantic-search parameters follow this pattern.

413 Batch Splitting

index_documents() in _operations/_patch.py recursively splits an IndexDocumentsBatch in half when the service returns RequestEntityTooLargeError (413). Retry-on-split uses is_retryable_status_code() (409/422/503).

SearchIndexingBufferedSender

Entirely hand-authored in _patch.py (sync) and aio/_patch.py (async); not generated. Wraps SearchClient with:

• Auto-flush timer (threading.Timer sync / asyncio.Task async).
• 413 recursive batch splitting.
• Retry per document key (409/422/503) bounded by max_retries_per_action.
• Key-field auto-detection against the index schema.
• The async version detects sync vs async callbacks via asyncio.iscoroutinefunction().

Field Builders and SearchFieldDataType Helpers (indexes/models/_patch.py)

• SimpleField(...) — sets searchable=False explicitly.
• SearchableField(...) — auto-types to String / Collection(String).
• ComplexField(...) — sets type to Complex / Collection(Complex).
• SearchField subclass adds hidden property (inverse of retrievable).
• SearchFieldDataType.Collection = staticmethod(_collection_helper) — monkey-patched onto the generated enum.

Backward-Compatible Enum Aliases

Applied at module load in indexes/models/_patch.py:

SearchFieldDataType.Int32 = SearchFieldDataType.INT32   # camelCase → UPPER
# (String, Int64, Single, Double, Boolean, DateTimeOffset, GeographyPoint, ComplexType)

After regeneration, verify every right-hand-side UPPER_CASE member still exists in the generated enum. If a new enum value collides with a Python keyword, add an alias.

Polymorphic Delete / Create-Or-Update Pattern

All delete/update operations in indexes/_operations/_patch.py accept str or model:

def delete_index(self, index, *, match_condition=MatchConditions.Unconditionally, **kwargs):
    try:
        name = index.name   # model object
        return self._delete_index(name=name, etag=index.e_tag, match_condition=match_condition, **kwargs)
    except AttributeError:
        return self._delete_index(name=index, **kwargs)   # string

Covered resources: index, synonym map, alias, indexer, data source connection, skillset, knowledge base, knowledge source. New resource types added to the spec need the same pair of wrappers in sync and async.

_convert_index_response Helper

list_indexes(select=...) returns SearchIndexResponse (projection type). _convert_index_response() maps it to SearchIndex — notably response.semantic → semantic_search. Shared between sync and async via import; do not duplicate.

When to use _patch.py vs TypeSpec: prefer TypeSpec decorators / emitter options for things the spec can express (renames, flattening, client names). Use _patch.py for Python-specific behavior: pagination contracts, threading/asyncio, polymorphic call patterns, 413 splitting, monkey-patched enum helpers, and backward-compat aliases. See typespec-python emitter docs.