Troubleshooting

Tokens Not Deanonymized in Response

Symptom: The LLM response contains raw tokens like <PERSON_1> instead of the original values.

Causes:

Non-streaming response with no session mappings. Deanonymization only runs on streaming responses when session.has_mappings is True. If the middleware didn't detect any entities in the request, no mappings exist and deanonymization is skipped.
Synthetic strategy mismatch. If the LLM invents its own names that happen to match synthetic values, those won't be in the session map. This is expected behavior — only values generated by the anonymizer are mapped.

Fix: Verify that entities are being detected by adding an on_entities_detected callback.

NER Model Not Found

Symptom: OSError: [E050] Can't find model 'en_core_web_sm'

Fix: Download the spaCy model:

python -m spacy download en_core_web_sm

And ensure you installed the NER extra:

pip install rdakt-ai[ner]

Redis Connection Refused

Symptom: ConnectionError: Error connecting to redis://localhost:6379

Fix: Ensure Redis is running and accessible at the configured URL:

redis-cli ping
# Should return: PONG

Import Error for RedisStore

Symptom: ImportError: redis package is required for RedisStore

Fix:

pip install rdakt-ai[redis]

Gzip Decompression Error

Symptom: zlib.error: Error -3 while decompressing data

Cause: This can happen if a proxy or CDN modifies the Content-Encoding header but doesn't actually decompress the body.

Fix: Ensure no middleware between your application and the Anthropic API is stripping or modifying encoding headers.

Custom Pattern Not Matching

Symptom: Your custom regex pattern doesn't detect entities.

Debug:

import re

pattern = r"EMP-\d{6}"
text = "Contact EMP-123456"
matches = re.finditer(pattern, text)
print(list(matches))

Common issues:

YAML requires escaping backslashes: "EMP-\\d{6}" not "EMP-\d{6}"
Python raw strings don't: r"EMP-\d{6}"

Streaming Latency

Symptom: Streaming responses feel delayed.

Cause: The deanonymizer buffers text when it encounters a < character that could be the start of a token. If the next chunk takes time to arrive, there's a perceived delay.

Mitigation: This is by design and only affects chunks containing potential token starts. The buffer has a 256-character safety limit. In practice, the delay is imperceptible (under 1ms) because LLM streaming chunks arrive rapidly.