Managed Ingestion Service

Managed Ingestion Service

Overview

Ingest documents into Amazon OpenSearch at scale via OSIS pipelines. Supports both Amazon OpenSearch Service (AOS) domains and Amazon OpenSearch Serverless (AOSS) collections.

When to Use

• User has an OpenSearch domain or AOSS collection and wants to ingest documents at scale
• User wants to move from local development to cloud ingestion
• User mentions OSIS, S3 ingestion, cloud processing, managed ingestion

Ingestion Paths

Both paths support the same sink options — the only difference is source content and processor.

Key Constraints

• AWS credentials must be configured; check first and stop if missing
• AOS domains: must be running OpenSearch 2.19+ for ASE; public domains only (no VPC)
• AOSS collections: must be in ACTIVE state before creating the pipeline
• The IAM role (sts_role_arn) must have S3 read + OpenSearch write permissions
• When using semantic_enrichment: it creates the index automatically — do NOT create it manually
• parse_json is the OSIS processor (parses JSONL lines); semantic_enrichment is in the sink (applies embeddings when writing) — they are independent
• Save generated YAML to .opensearch/pipelines/<pipeline-name>.yaml

Entry Point

Step 1: Validate credentials

aws sts get-caller-identity

If this fails, stop and ask the user to configure credentials.

Step 2: Determine the deploy target

Ask: "Are you ingesting into an Amazon OpenSearch Service (AOS) domain or an OpenSearch Serverless (AOSS) collection?"

AOS domain:

• Must be running OpenSearch 2.19+ for semantic_enrichment
• Verify the domain exists and is in ACTIVE state:

  aws opensearch describe-domain --domain-name <domain-name> --region <region>
  

AOSS collection:

• Verify it's ACTIVE:

  aws opensearchserverless batch-get-collection \
    --ids <collection-id> --region <region>
  

If the user needs a new domain/collection, provision one using the aws-setup skill. Wait until ACTIVE, then continue here.

Step 3: Determine the data source

Ask: "Do you already have JSONL chunks, or do you have raw documents (PDF/DOCX) that need processing?"

If the user already has JSONL chunks:

• Validate the JSONL format: each line must be valid JSON with text fields that can be enriched.
• Proceed to Step 4.

If the user has raw documents (PDF/DOCX/etc.):

Present both options to the user:

I see you have raw PDF documents. There are two ways to process them for ingestion:

Option A — Cloud ingestion (recommended for scale) Upload your PDFs directly to S3 and let the cloud handle parsing, chunking, and indexing with ASE (automatic sparse embeddings) — no local compute needed. This is currently in private beta.

👉 To get access, email opensearch-agent-skills-interests@amazon.com with your AWS account ID and region.

Option B — Local processing (available now) Process your PDFs locally with Docling to produce JSONL chunks, then upload those to S3 for cloud ingestion. Gives you full control over chunk quality.

Which would you prefer? (Or if you'd like to get started immediately, I'll go with Option B while you wait for beta access.)

• If the user chooses Option A: confirm they should email opensearch-agent-skills-interests@amazon.com, and offer to proceed with Option B in the meantime.
• If the user chooses Option B: proceed with document_processing_guide.md. Once chunks are ready at .opensearch/chunks/<index>/, return here and continue to Step 4.

Step 4: Gather parameters

Collect from the user (one at a time):

1. Deploy target — AOS domain endpoint or AOSS collection endpoint
2. S3 bucket name — where chunks will be uploaded
3. S3 prefix — key prefix for source files (e.g., chunks/input)
4. Target index name — will be created by semantic_enrichment if new; or an existing index
5. IAM role ARN — for OSIS pipeline (or offer to create one)
6. Network policy name — (AOSS only) network policy allowing OSIS access
7. Region — AWS region

Step 5: Determine the sink mode

Ask: "Do you already have an existing index to ingest into, or should I create a new one with automatic semantic enrichment (ASE)?"

Path A — Existing index (no semantic_enrichment):

• The user already has an index with mappings configured.
• Use the plain opensearch sink (no semantic_enrichment block).
• Documents are written directly without automatic embedding.
• Proceed with Pipeline YAML — Existing Index.

Path B — New index (with semantic_enrichment / ASE):

• OSIS creates the index automatically with ASE.
• Ask: "Which text fields in your chunks should have semantic enrichment (sparse embeddings)?"
• The user should specify the field names from their JSONL that contain natural language text (e.g., text, title, description, summary).
• Each field is enriched independently — ASE generates sparse vectors for each specified field.
• Proceed with Pipeline YAML — New Index with Semantic Enrichment.

Note: semantic_enrichment is about which text fields get sparse embeddings — it is not tied to any particular "search strategy." Any text field that contains natural language and should be semantically searchable should be listed.

Step 6: Upload to S3

Upload your locally-prepared chunks:

aws s3 cp .opensearch/chunks/<index>/<file>.jsonl s3://<bucket>/<prefix>/

For multiple files:

aws s3 cp .opensearch/chunks/<index>/ s3://<bucket>/<prefix>/ --recursive

Step 7: Create OSIS pipeline

Generate the YAML based on the selected path and sink mode, save to .opensearch/pipelines/<pipeline-name>.yaml, then create:

aws osis create-pipeline \
  --pipeline-name <pipeline-name> \
  --min-units 1 \
  --max-units 1 \
  --pipeline-configuration-body file://.opensearch/pipelines/<pipeline-name>.yaml \
  --region <region>

Step 8: Verify

Poll until ACTIVE:

aws osis get-pipeline \
  --pipeline-name <pipeline-name> \
  --region <region> \
  --query 'Pipeline.Status' --output text

Then verify documents are indexed (check doc count, test a match query — if ASE is configured, it rewrites match queries to neural sparse automatically).

Pipeline YAML — Local Chunking + New Index (AOSS with semantic_enrichment)

version: "2"
<pipeline-name>:
  source:
    s3:
      codec:
        newline:
      compression: "none"
      aws:
        region: "<region>"
        sts_role_arn: "<iam-role-arn>"
      acknowledgments: true
      scan:
        scheduling:
          interval: PT30S
        buckets:
          - bucket:
              name: "<s3-bucket-name>"
              filter:
                include_prefix:
                  - <s3-prefix>
  processor:
    - parse_json:
    - delete_entries:
        with_keys: [ "s3" ]
  sink:
    - opensearch:
        hosts: [ "<aoss-collection-endpoint>" ]
        aws:
          region: "<region>"
          sts_role_arn: "<iam-role-arn>"
          semantic_enrichment:
            fields:
              - name: "<text-field-name>"
                language: "english"
          serverless: true
          serverless_options:
            network_policy_name: "<network-policy-name>"
        index: "<index-name>"
        index_type: custom

Pipeline YAML — Local Chunking + New Index (AOS domain with semantic_enrichment)

version: "2"
<pipeline-name>:
  source:
    s3:
      codec:
        newline:
      compression: "none"
      aws:
        region: "<region>"
        sts_role_arn: "<iam-role-arn>"
      acknowledgments: true
      scan:
        scheduling:
          interval: PT30S
        buckets:
          - bucket:
              name: "<s3-bucket-name>"
              filter:
                include_prefix:
                  - <s3-prefix>
  processor:
    - parse_json:
    - delete_entries:
        with_keys: [ "s3" ]
  sink:
    - opensearch:
        hosts: [ "<aos-domain-endpoint>" ]
        aws:
          region: "<region>"
          sts_role_arn: "<iam-role-arn>"
          semantic_enrichment:
            fields:
              - name: "<text-field-name>"
                language: "english"
        index: "<index-name>"
        index_type: custom

Key difference from AOSS: no serverless: true or serverless_options block.

Pipeline YAML — Existing Index (no semantic_enrichment)

Use this when the user already has an index with mappings configured.

For AOSS:

version: "2"
<pipeline-name>:
  source:
    s3:
      codec:
        newline:
      compression: "none"
      aws:
        region: "<region>"
        sts_role_arn: "<iam-role-arn>"
      acknowledgments: true
      scan:
        scheduling:
          interval: PT30S
        buckets:
          - bucket:
              name: "<s3-bucket-name>"
              filter:
                include_prefix:
                  - <s3-prefix>
  processor:
    - parse_json:
    - delete_entries:
        with_keys: [ "s3" ]
  sink:
    - opensearch:
        hosts: [ "<aoss-collection-endpoint>" ]
        aws:
          region: "<region>"
          sts_role_arn: "<iam-role-arn>"
          serverless: true
          serverless_options:
            network_policy_name: "<network-policy-name>"
        index: "<index-name>"
        index_type: custom

For AOS:

version: "2"
<pipeline-name>:
  source:
    s3:
      codec:
        newline:
      compression: "none"
      aws:
        region: "<region>"
        sts_role_arn: "<iam-role-arn>"
      acknowledgments: true
      scan:
        scheduling:
          interval: PT30S
        buckets:
          - bucket:
              name: "<s3-bucket-name>"
              filter:
                include_prefix:
                  - <s3-prefix>
  processor:
    - parse_json:
    - delete_entries:
        with_keys: [ "s3" ]
  sink:
    - opensearch:
        hosts: [ "<aos-domain-endpoint>" ]
        aws:
          region: "<region>"
          sts_role_arn: "<iam-role-arn>"
        index: "<index-name>"
        index_type: custom

Key points:

• No semantic_enrichment — the index already exists with its own mappings
• Documents are written directly without automatic embedding
• Use this when the user manages their own index lifecycle

Cloud Ingestion (private beta)

Private beta. Cloud ingestion lets you upload raw documents (PDF/DOCX) directly to S3 and have the OSIS pipeline handle parsing, chunking, and indexing end-to-end — no local compute needed. Contact opensearch-agent-skills-interests@amazon.com to onboard. Pipeline YAML will be provided during onboarding.

Key Concepts

parse_json vs semantic_enrichment

These are independent components at different stages of the pipeline:

You can use parse_json without semantic_enrichment (existing index) or with it (new index + ASE).

semantic_enrichment field selection

semantic_enrichment is about which text fields get sparse vector embeddings. It is NOT tied to any search strategy (hybrid, semantic, agentic). Any text field containing natural language that you want to be semantically searchable should be listed:

semantic_enrichment:
  fields:
    - name: "text"
      language: "english"
    - name: "title"
      language: "english"

Ask the user: "Which fields in your chunks contain natural language text that should be semantically searchable?" — then list those fields.

IAM Role Setup (if needed)

See iam-setup.md for full IAM role creation, S3/AOSS/AOS permissions, and data access policy configuration. The IAM role is a prerequisite — without it the pipeline will fail with permission errors on every retry.

Parameters Reference