Compact

# Compact Filter Example
# Requires `--features openai-file-resolve-filter,openai-compact,store-sqlite` because these filters are opt-in.
#
# Demonstrates compaction after rehydrate, file resolve, and document
# extract so rewritten current-turn content survives history
# replacement. Store a response, rehydrate it on the next turn, and
# summarize when the token count exceeds the compact threshold.
#
# Pipeline order is required: file_resolve and doc_extract rewrite the
# current-turn tail in place and leave `state.input` as the original
# client payload. Compact preserves that rewritten tail rather than
# restoring `file_url` or `input_file`.
#
# Security: StreamBuffer body callouts run before this listener's
# header-phase filters. Deploy this example behind an outer
# authentication and authorization boundary before enabling the
# required allow_pre_security_callout acknowledgement below.
#
# Usage:
#   cargo run -p praxis-ai-proxy --features openai-file-resolve-filter,openai-compact,store-sqlite -- \
#     -c examples/configs/openai/responses/compact.yaml
#
# Requires an OpenAI-compatible backend on port 11434 (e.g. Ollama).
#
# Test:
#   # Step 1 — store a response
#   RESP_ID=$(curl -s -X POST http://localhost:8080/v1/responses \
#     -H "Content-Type: application/json" \
#     -d '{"model":"llama3.2:1b","input":"Explain TCP vs UDP in detail"}' \
#     | python3 -c "import sys,json; print(json.load(sys.stdin)['id'])")
#
#   # Step 2 — follow up with compaction (threshold=1000 tokens)
#   curl -s -X POST http://localhost:8080/v1/responses \
#     -H "Content-Type: application/json" \
#     -d "{\"model\":\"llama3.2:1b\",\"input\":\"Compare with QUIC\",\"previous_response_id\":\"$RESP_ID\",\"context_management\":[{\"type\":\"compaction\",\"compact_threshold\":1000}],\"store\":false}"
#
#   # Step 3 — compact a follow-up that includes a file_url (needs a
#   # Files API on 127.0.0.1:9999 only for file_id; file_url is fetched
#   # directly). The upstream body must contain resolved file_data or
#   # extracted input_text, never the original file_url.

listeners:
  - name: ai-gateway
    address: "127.0.0.1:8080"
    filter_chains: [compact-pipeline]

filter_chains:
  - name: compact-pipeline
    filters:
      - filter: openai_responses_request
        on_invalid: continue
        headers:
          format: x-praxis-ai-format
          model: x-praxis-ai-model
          stream: x-praxis-ai-stream
          mode: x-praxis-responses-mode

      - filter: openai_tool_parse

      - filter: state_owner
        mode: single_tenant
        tenant_id: default

      - filter: openai_response_store
        backend: sqlite
        database_url: "sqlite://responses.db?mode=rwc"
        responses_table: openai_responses
        conversations_table: openai_conversations

      # `openai_stream_events` is intentionally absent: it composes streamed
      # inference rounds into one logical Responses stream and is only valid
      # inside an `iterative_request_router` step. This compaction demo proxies
      # finite responses, so it would be inert here. See `stream-events.yaml`
      # and `agentic-loop.yaml` for the in-IRR streaming examples.
      - filter: openai_responses_rehydrate

      - filter: openai_file_resolve
        files_api_url: "http://127.0.0.1:9999"
        allow_pre_security_callout: true
        # file_id callouts run through this outbound filter chain via the
        # filtered subrequest executor. SSRF protection for files_api_url
        # derives from the pipeline's allow_private_upstreams; client
        # file_url downloads never traverse this chain.
        outbound_chain:
          name: files-api-outbound
          filters:
            - filter: headers
              request_set:
                - name: x-file-callout
                  value: file-resolve
        file_url: resolve
        forward_headers:
          - authorization
        on_missing: reject
        timeout_ms: 10000

      - filter: openai_doc_extract
        allow_pre_security_callout: true
        on_unsupported: continue

      - filter: openai_responses_compact
        # This direct summarization callout is anonymous; no downstream or
        # cluster authentication headers are forwarded.
        allow_pre_security_callout: true
        inference_url: "http://localhost:11434/v1/chat/completions"
        allow_private_inference_url: true
        default_model: llama3.2:1b
        timeout_ms: 60000

      - filter: openai_responses_proxy
        name: inference

      - filter: router
        routes:
          - path: "/v1/responses"
            headers:
              x-praxis-ai-format: "openai_responses"
            cluster: "inference-backend"

      - filter: load_balancer
        clusters:
          - name: "inference-backend"
            endpoints:
              - "127.0.0.1:11434"

insecure_options:
  allow_private_endpoints: true # example proxies to local backends
  allow_private_upstreams: true # file_id callouts reach the loopback Files API