embedder: guard against long inputs by chunking, batching and averaging

[Yuyuyaka]

Title: embedder: guard against long inputs by chunking, batching and averaging

Summary

This patch makes the OpenAI-based embedder resilient to very long text inputs
that previously triggered provider token-limit errors (causing 500s). It adds
character-based chunking, batching and averaging of per-chunk embeddings as a
safe default, and includes a minimal Dockerfile so maintainers can reproduce
and publish a patched image.

What changed

• OpenAIEmbedder.create now:
  • chunks long str inputs into roughly DEFAULT_MAX_INPUT_CHARS characters,
  • embeds chunks in batches and averages resulting vectors into a single
    representation,
  • falls back to a direct call if chunked embedding fails.
• create_batch delegates to create() so each item is guarded.
• Added docker/Dockerfile + build_image.sh for a reproducible local image.

Why

Long messages (> provider token limit) caused the upstream embedding API to
reject requests and Graphiti to return 500, which cascaded to OpenClaw gateway
timeouts. This change prevents oversized inputs from being forwarded directly
to the provider and provides a conservative, backwards-compatible default.

Testing done

• Verified locally inside the running Graphiti container by copying the
  patched openai.py and restarting: POST /get-memory with ~20k chars now
  returns HTTP 200 and does not raise embedding errors.
• Built a reproducible image locally via docker build (tag
  graphiti-local:embedder-chunking-20260512).

How to review

• Review the new openai.py implementation for correctness and style.
• Consider whether averaging embeddings is the desired aggregation strategy or
  whether the code should instead store multiple per-chunk embeddings and
  change retrieval logic.

Notes for maintainers

• This is a conservative, defensive change. Alternatives: per-chunk storage,
  more sophisticated summarization before embedding, or token-aware chunking.
• If accepted, please publish a patch release and update downstream images.