Bug: embedding fails with 400 on long content (no chunking or size handling)

[johnathanneals-dev]

Summary

create_embedding() passes the full text to the embedding provider with no size handling. When the input exceeds the model's context window (e.g., 2048 tokens for nomic-embed-text, 8191 for text-embedding-3-small), the API returns a 400 error and the memory store operation fails silently or crashes.

This affects any memory that exceeds the model's token limit — long conversation transcripts, detailed technical notes, multi-section documents.

Reproduction

from src.embedder import create_embedding

long_text = "detailed technical content... " * 2000  # ~10KB
embedding = create_embedding(long_text)  # 400 error from provider

With nomic-embed-text (Ollama, 2048-token context):

requests.exceptions.HTTPError: 400 Client Error: {"error":"the input length exceeds the context length"}

With OpenRouter/OpenAI models the threshold is higher (~8K tokens) but the same crash occurs on sufficiently long content.

Root cause

src/embedder/__init__.py — the embed() method on every provider class sends the full input text without checking length or chunking:

def embed(self, text: str) -> List[float]:
    response = requests.post(
        f'{self.base_url}/embeddings',
        json={'model': self.model, 'input': text},
        ...
    )

No provider implementation has size checking, truncation, or chunking.

Suggested fix

Add automatic chunking with mean-pooling at the create_embedding() level (above the provider abstraction, so all providers benefit):

1. Split text exceeding max_chars into overlapping chunks
2. Embed each chunk independently
3. Mean-pool + L2-normalize the chunk embeddings into a single vector

This preserves the single-vector-per-memory contract while handling arbitrarily long input. The full text is still stored in the database — only the embedding generation is affected.

A conservative default of ~4000 chars safely fits within nomic-embed-text's 2048-token window even for dense technical content (which tokenizes at ~2-3 chars/token). Higher-capacity models like text-embedding-3-small could use a larger threshold via configuration.