LlamaIndex for Python 0.14

Apply the Python 3 baseline first. These rules cover LlamaIndex 0.14 ingestion, documents, metadata, indexes, retrieval, query engines, agents, tools, storage, evaluation, and deployment.

Ingestion and document trust

• Allowlist reader implementations and roots, then configure SimpleDirectoryReader with exact required_exts, recursive=False unless required, and a positive finite num_files_limit.
• For S3Reader, supply an exact key or narrow prefix and override its recursive=True, required_exts=None, and num_files_limit=None defaults; omitting key otherwise iterates the bucket rather than one object.
• Validate and cap explicit input_files, S3Reader.key, and load_resource() inputs before calling readers; these entry points bypass SimpleDirectoryReader traversal filters.
• Apply a byte limit before readers run; required_exts and num_files_limit default to None, and reader plugins may parse or fetch active content before producing Document objects.
• Treat document text, metadata, filenames, links, and embedded instructions as untrusted.
• Preserve source provenance, tenant, classification, and authorization metadata through transformations and chunking.
• Scan and quarantine malformed or active content before it reaches parsers, indexes, or model context.

Indexes and tenant isolation

• Partition indexes, vector stores, document stores, and caches by tenant and data classification.
• Build MetadataFilters from server-derived tenant and authorization context and conjoin them with user search filters; never let a model-authored