[Bug]: Memory commit triggers oversized embedding input → unhandled exception hangs uvicorn

[laofahai]

Summary

After memory extraction via session.commit(), the semantic processor generates .overview.md for the parent memory directory. When this overview text exceeds the embedding model's context length, OpenAIDenseEmbedder.embed() raises an unhandled RuntimeError. This exception appears to block the uvicorn event loop, causing the entire HTTP server to become unresponsive (process alive, port open, but all endpoints hang).

Environment

• OpenViking: installed via pipx (latest as of 2026-03-17)
• OS: macOS arm64 (Darwin 25.3.0)
• Embedding: Ollama nomic-embed-text (8192 token context, 768 dim) via OpenAI-compatible API
• VLM: Bailian qwen3-max
• Mode: local, bound to 127.0.0.1:1933
• Integration: OpenClaw memory-openviking plugin

Steps to Reproduce

1. Accumulate enough memories in a directory (e.g. viking://user/default/memories/preferences)
2. Trigger a new memory extraction (e.g. via session commit from OpenClaw)
3. Memory extractor writes the new memory file successfully
4. Semantic processor runs on the parent directory (recursive=False)
5. Generated .overview.md aggregates all file summaries → text exceeds embedding model's token limit
6. Embedding queue calls OpenAIDenseEmbedder.embed() with the oversized text
7. Ollama returns HTTP 400: the input length exceeds the context length
8. embed() raises RuntimeError → collection_schemas.py:on_dequeue propagates the exception
9. Server hangs: all HTTP endpoints stop responding, curl times out

Relevant Logs

INFO - Processing semantic generation for: viking://user/default/memories/preferences (recursive=False)
WARNING - Candidate data is None for label index 4 (label: ...), skipping.
INFO - Created memory file: viking://user/default/memories/preferences/mem_04c2ef28-...md
INFO - Enqueued memory for vectorization

stderr:

openai.BadRequestError: Error code: 400 - {'error': {'message': 'the input length exceeds the context length', ...}}
RuntimeError: OpenAI API error: Error code: 400 - ...

After this error, no further log output appears and all HTTP requests time out.

Root Cause Analysis

Two issues combine:

1. No input truncation guard in OpenAIDenseEmbedder.embed() (openai_embedders.py): text is passed directly to the API without any length check. When the embedding model has a limited context window, oversized input causes a hard API error.

2. Unhandled exception in embedding queue blocks uvicorn: the RuntimeError from the embedder propagates through collection_schemas.py:on_dequeue and appears to block or crash the async event loop, making the entire server unresponsive.

Expected Behavior

• Embedding input should be truncated (or chunked) before being sent to the provider
• Embedding failures should be caught gracefully without blocking the HTTP server
• The server should remain responsive even if individual vectorization tasks fail

Workaround

Monkey-patched OpenAIDenseEmbedder.embed() to truncate input to 24000 chars (~6000-8000 tokens) before calling the API. Server remains stable after the patch.

Related Issues

• [Bug]: add-resource sends oversized input to OpenAI embeddings API during repo import #616 — same oversized embedding problem but for add-resource (closed, only addressed that path)
• Long memory indexing should use first-class chunked vectorization instead of relying on single-record embedding #530 — long memory should use chunked vectorization (design discussion)
• Embedding truncation and chunking should have clearer responsibilities across memory, file, and directory vectorization #531 — truncation vs chunking responsibilities (design discussion)
• Bug: process alive but HTTP endpoints hang, causing AbortError in memory_store/auto-capture #527 — process alive but HTTP endpoints hang (same symptom, likely same root cause)
• Memory extraction triggers O(n²) semantic reprocessing — token cost grows quadratically with memory count #505 — O(n²) semantic reprocessing amplifies the problem

[qin-ctx]

Thanks for the excellent bug report! The reproduction steps, logs, root cause analysis, and workaround are all very thorough — this is incredibly helpful for tracking down and fixing the issue.

Your analysis is spot on — there are two problems that need to be addressed:

1. OpenAIDenseEmbedder.embed() lacks an input length guard
2. Unhandled exceptions in the embedding queue can block the event loop and hang the server

#616 previously fixed the add-resource path, but this memory commit path was missed.

Would you be interested in contributing a PR to fix this? If so, we'd be happy to provide guidance!

[laofahai]

Thanks for the quick response! After investigating, I found that both issues are already fixed in v0.2.6:

1. Input length guard: OpenAIDenseEmbedder.embed() now checks _estimate_tokens(text) > max_tokens and calls _chunk_and_embed() for oversized input (commit cbf8bef, PR fix(embed): add text chunking for oversized embedding inputs (#616) #642).

2. Exception handling in on_dequeue: The embedding call is now wrapped in an inner try/except (line 248-271 in collection_schemas.py), so failures are caught gracefully via report_error() without blocking the event loop.

I was running v0.2.5 where both were missing. Upgraded to v0.2.6 via pipx upgrade openviking and confirmed the fix — embedding works correctly with automatic chunking, and the server stays responsive.

This issue can be closed. Thanks!