Text, images, audio and tables do not become suitable foundation-model inputs merely because they have been copied into an S3 bucket. An engineering team must validate each input’s ownership, quality, structure, supported format and business meaning before inference. In the AWS Certified Generative AI Developer \u2013 Professional (AIP-C01) blueprint, processing heterogeneous data is part of designing production foundation-model integration, not an optional pre-processing detail.<\/p>\n
Imagine an insurance intake system that receives a photographed form, a scanned PDF policy, a call recording and a spreadsheet of covered limits. If the photographed form cuts off a date, the audio transcription mishears a claim number and the spreadsheet extraction loses column headers, the model might produce a coherent but incorrect coverage summary. The safest system preserves source provenance, identifies uncertain extraction and prevents unsupported conclusions from becoming decisions.<\/p>\n
Before building a common pipeline, collect the input’s MIME type, extension, size, origin, account scope, consent status, retention obligation and required model capability. Do not use a file extension as a security boundary; validate actual content types and reject disallowed or suspicious payloads. Treat uploaded files and extracted text as untrusted data that might contain instructions designed to redirect an agent.<\/p>\n
Specify what the output must represent. A scanned table might require exact cells and units rather than a general summary. A customer call may need a speaker-attributed transcript before an agent can extract commitments. A product photo might need evidence of visible damage rather than an unrestricted description. A model response cannot recover pixels, words or column relationships lost earlier in the pipeline.<\/p>\n
For structured HTML, PDF text or office documents, extract a usable text representation while preserving headings, lists, tables, page references and effective dates. Strip unneeded markup but retain source offsets or passage IDs that support citations. Compare extracted text against known golden files, especially where page layout affects meaning: a negative qualifier in a separate table cell can reverse an entitlement.<\/p>\n
Normalize character encodings and whitespace, resolve language or locale where necessary, and detect truncated or empty extracts. Preserve the original object securely for authorized audit. Use a review or reprocessing queue for records whose quality checks fail instead of including meaningless text in a vector index. Amazon Bedrock Knowledge Bases<\/a> can support retrieval pipelines, but document validity and access classification remain the organization’s responsibility.<\/p>\n For images, check format support, pixel dimensions, orientation, corruption, possible sensitive content and sufficient legibility for the intended question. A diagram or screenshot can carry labels whose spatial relationship matters; a text-only OCR output may discard that relationship. Depending on the supported model and workflow, consider a multimodal representation or a structured extraction step, then evaluate the result against the source image.<\/p>\n For audio, inspect length, sample rate, language, silence, clipping, overlapping speakers and consent. A low-confidence transcription should not be converted directly into a legal admission. Use a suitable transcription service, such as Amazon Transcribe where appropriate, or supported multimodal processing, and keep confidence and speaker-attribution limitations visible to downstream reviewers. A clean transcript can still lack speaker identity or contain a crucial homophone error.<\/p>\n AWS’s multimodal Knowledge Bases approach guidance<\/a> distinguishes alternatives based on modality and task. Choose supported processing paths rather than assuming image, audio, video and documents have interchangeable model APIs or accepted file formats. The exact feature availability may vary by Region, model and service configuration.<\/p>\n A spreadsheet containing premiums or dosage-like quantities requires extra care about units, column identity, null values and calculation provenance. Parse the workbook or source table deterministically where possible. Validate schema and referential integrity before converting it to prompt context. If the application expects a monetary limit, include currency and effective policy version rather than letting a model infer missing context from similar rows.<\/p>\n Extracted entity values should be linked to their original source fields. A contract amount may appear in the document header, footnotes and an amended schedule; choose the authoritative value under explicit business rules. If conflicting records cannot be reconciled, return the conflict for human review. Use automatic quality rules and exception handling rather than asking the model to guess which entry is current.<\/p>\nBuild different quality gates for images and audio<\/h2>\n
Treat tables as records, not prose with numbers<\/h2>\n
Prepare inference payloads for a specific supported model<\/h2>\n