Skip to content

DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.

fix(llm): avoid tokenizing image payloads in request fingerprints (#1712)

ZyPulse07948c038574
diff --git a/deeptutor/services/llm/request_cache.py b/deeptutor/services/llm/request_cache.py
index a55f2502b..91053aa25 100644
--- a/deeptutor/services/llm/request_cache.py
+++ b/deeptutor/services/llm/request_cache.py
@@ -24,16 +24,17 @@ def fingerprint_request(
tools: Any,
route: dict[str, Any],
) -> dict[str, Any]:
- from deeptutor.services.session.context_builder import count_tokens
+ from deeptutor.services.session.context_builder import _count_model_context_tokens
- # Hash an entire message, never individual tokens (which disclose vocabulary).
+ # Hash the complete replay payload, including image bytes. Only the token
+ # estimate uses the image-aware view; Base64 and URLs are not language tokens.
return {
"route": _digest(route),
"tools": _digest(tools or []),
"messages": [
{
"hash": _digest(message),
- "tokens": count_tokens(_encoded(message)),
+ "tokens": _count_model_context_tokens(message),
"role": message.get("role"),
}
for message in messages
diff --git a/tests/services/llm/test_request_cache.py b/tests/services/llm/test_request_cache.py
index 87908928d..18c2f66a4 100644
--- a/tests/services/llm/test_request_cache.py
+++ b/tests/services/llm/test_request_cache.py
@@ -63,3 +63,67 @@ def test_claude_compat_marks_latest_tool_result_with_at_most_four_breakpoints():
)
assert count <= 4
assert (messages, tools) == original
+
+
+def test_image_fingerprint_keeps_payload_hash_without_tokenizing_image_data(monkeypatch):
+ from copy import deepcopy
+
+ from deeptutor.services.session import context_builder
+
+ payload = "YWJj" * 1000
+ blocks = [
+ {"type": "image_url", "image_url": {"url": "data:image/png;base64," + payload}},
+ {"type": "input_image", "image_url": "https://example.org/image.png"},
+ {"type": "input_image", "file_id": "file-example"},
+ {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": payload}},
+ ]
+ messages = [{"role": "user", "content": [*blocks, {"type": "text", "text": "hello"}]}]
+ original = deepcopy(messages)
+ tokenized = []
+
+ def count(text):
+ tokenized.append(text)
+ return 7
+
+ monkeypatch.setattr(context_builder, "count_tokens", count)
+ fingerprint = fingerprint_request(messages, [], {})
+ assert payload not in "".join(tokenized)
+ assert "https://example.org" not in "".join(tokenized)
+ assert "file-example" not in "".join(tokenized)
+ assert "hello" in "".join(tokenized)
+ assert (
+ fingerprint["messages"][0]["tokens"] == 7 + 4 * context_builder.IMAGE_CONTEXT_TOKEN_ESTIMATE
+ )
+ assert messages == original
+ changed = deepcopy(messages)
+ changed[0]["content"][0]["image_url"]["url"] += "AAAA"
+ assert (
+ fingerprint_request(changed, [], {})["messages"][0]["hash"]
+ != fingerprint["messages"][0]["hash"]
+ )
+ assert (
+ compare_requests(fingerprint, fingerprint_request(changed, [], {}))["first_change"]
+ == "history"
+ )
+
+
+def test_fingerprint_estimate_preserves_tool_and_provider_replay_text(monkeypatch):
+ from deeptutor.services.session import context_builder
+
+ tokenized = []
+ monkeypatch.setattr(context_builder, "count_tokens", lambda text: tokenized.append(text) or 11)
+ messages = [
+ {
+ "role": "assistant",
+ "content": "reasoning",
+ "tool_calls": [{"arguments": "important"}],
+ "thinking_blocks": [{"signature": "signature", "text": "thought"}],
+ },
+ {"role": "tool", "content": "result", "tool_call_id": "call"},
+ ]
+ fingerprint = fingerprint_request(messages, [], {})
+ assert [message["tokens"] for message in fingerprint["messages"]] == [11, 11]
+ assert all(
+ value in "".join(tokenized)
+ for value in ("reasoning", "important", "signature", "thought", "result", "call")
+ )