fix: limit resolve_extracted_nodes context to prevent max_tokens overflow

[rafaelreis-r]

Problem

When deduplicating extracted nodes in resolve_extracted_nodes, the LLM prompt includes the full candidate.attributes dict for every candidate returned by the similarity search. In production workloads with:

• 10–15 extracted nodes per episode
• Up to 10 candidates per node from vector/BM25 search

This produces ~100–150 candidates, each carrying verbose attributes (summaries, descriptions, relationship lists, etc.). The resulting prompt regularly exceeds 17k tokens, pushing the JSON output past the max_tokens=16384 limit and causing truncated/invalid responses.

Reported in issue #1275.

Fix

Two surgical changes to graphiti_core/utils/maintenance/node_operations.py:

1. Remove candidate.attributes from the resolution context

Only name and entity_types are needed for identity-level deduplication. The full attribute payload adds thousands of tokens without improving match quality.

# Before
existing_nodes_context = [
    {
        **{'name': candidate.name, 'entity_types': candidate.labels},
        **candidate.attributes,   # ← verbose, causes overflow
    }
    for candidate in indexes.existing_nodes
]

# After
existing_nodes_context = [
    {
        'name': candidate.name,
        'entity_types': candidate.labels,
        # attributes omitted to prevent token overflow
    }
    for candidate in indexes.existing_nodes[:MAX_RESOLVE_CANDIDATES]
]

2. Add MAX_RESOLVE_CANDIDATES = 50 constant

Caps the candidate list to bound worst-case context size regardless of search result count. Placed alongside the existing MAX_NODES constant.

Impact

• No change to resolution logic or search behavior
• Eliminates context overflow for typical workloads (10–15 nodes × 10 candidates)
• MAX_RESOLVE_CANDIDATES = 50 provides a safety ceiling for extreme cases
• The name + entity_types fields are sufficient for the deduplication LLM to make correct identity judgments

[danielchalef]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

Rafael Reis seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}