Chat Completions To Converse

# OpenAI Chat Completions to AWS Bedrock Converse
#
# Accepts OpenAI Chat Completions requests and transparently forwards
# them to AWS Bedrock Converse, translating both the request and
# response bodies on the fly.
#
# Pipeline overview
# ─────────────────
# 1. `openai_chat_completions_to_bedrock_converse` — Praxis pre-reads the
#    full request body before processing the filter chain. During pre-read,
#    this filter translates the Chat Completions payload to the Bedrock
#    Converse shape and sets the upstream path to:
#      /model/{model-id}/converse       (non-streaming)
#      /model/{model-id}/converse-stream (streaming)
#    Both the rewritten path and the translated body are available to
#    subsequent filters during their request-header phase.
#
# 2. `router` — routes on the rewritten /model/{id}/... path prefix,
#    selecting the Bedrock Runtime cluster.
#
# 3. `load_balancer` — forwards requests to the Bedrock Runtime endpoint.
#
# 4. `aws_sigv4_sign` — signs the upstream request with AWS Signature
#    Version 4. Because Praxis delivers the translated body and rewritten
#    path to this filter's request-header phase (see step 1 above), the
#    signature covers the correct Bedrock path and body hash. It also
#    injects the Host header and strips the client-supplied Authorization.
#
# Non-streaming and streaming responses are both handled by the
# translation filter:
#   - Non-streaming: the full Bedrock Converse JSON is buffered and
#     rewritten to a Chat Completions response.
#   - Streaming: each binary AWS EventStream frame is decoded and
#     emitted as an OpenAI SSE `data: {chunk}\n\n` event. The stream
#     ends with `data: [DONE]\n\n`.
#
# Required environment variables:
#   AWS_ACCESS_KEY_ID
#   AWS_SECRET_ACCESS_KEY
#
# In this example the backend cluster points at a local test double
# (127.0.0.1:3000). In production, replace the endpoint with the real
# Bedrock Runtime regional hostname and enable TLS:
#
#   endpoints:
#     - "bedrock-runtime.us-east-1.amazonaws.com:443"
#   tls:
#     sni: bedrock-runtime.us-east-1.amazonaws.com
#
# The `host` field in `aws_sigv4_sign` must match the cluster's actual
# TLS SNI/hostname — Praxis does not reconcile them automatically.

listeners:
  - name: gateway
    address: "0.0.0.0:8080"
    filter_chains:
      - bedrock-translate

filter_chains:
  - name: bedrock-translate
    filters:
      # Step 1 — Translate Chat Completions ↔ Bedrock Converse.
      # This filter pre-reads the full request body; the translated body
      # and rewritten upstream path are ready before any subsequent
      # on_request callbacks run.
      - filter: openai_chat_completions_to_bedrock_converse
        max_body_bytes: 4194304        # optional; default 4 MiB

      # Step 2 — Route on the rewritten /model/… path.
      - filter: router
        routes:
          - path_prefix: "/model/"
            cluster: bedrock

      # Step 3 — Select the Bedrock Runtime endpoint.
      - filter: load_balancer
        clusters:
          - name: bedrock
            endpoints:
              - "127.0.0.1:3000"      # replace with bedrock-runtime.{region}.amazonaws.com:443

      # Step 4 — Sign the request with AWS SigV4.
      # Reads ctx.rewritten_path (/model/{id}/converse[-stream]) and
      # ctx.buffered_request_body (the translated Bedrock JSON) so the
      # signature covers the actual upstream path and body.
      - filter: aws_sigv4_sign
        region: us-east-1
        service: bedrock
        host: 127.0.0.1:3000             # replace with the production Bedrock Runtime hostname
        access_key_env_var: AWS_ACCESS_KEY_ID
        secret_key_env_var: AWS_SECRET_ACCESS_KEY

insecure_options:
  allow_private_endpoints: true # example proxies to a local test double