Fix missing json (#3177)

* initial steps

* k

* remove logs

* k

* k

backend/danswer/llm/answering/stream_processing/quotes_processing.py

@@ -231,16 +231,16 @@ class QuotesProcessor:
 
         model_previous = self.model_output
         self.model_output += token
-
         if not self.found_answer_start:
             m = answer_pattern.search(self.model_output)
             if m:
                 self.found_answer_start = True
 
                 # Prevent heavy cases of hallucinations
-                if self.is_json_prompt and len(self.model_output) > 70:
-                    logger.warning("LLM did not produce json as prompted")
+                if self.is_json_prompt and len(self.model_output) > 400:
                     self.found_answer_end = True
+                    logger.warning("LLM did not produce json as prompted")
+                    logger.debug("Model output thus far:", self.model_output)
                     return
 
                 remaining = self.model_output[m.end() :]

backend/danswer/utils/text_processing.py

@@ -4,6 +4,10 @@ import re
 import string
 from urllib.parse import quote
 
+from danswer.utils.logger import setup_logger
+
+
+logger = setup_logger(__name__)
 
 ESCAPE_SEQUENCE_RE = re.compile(
     r"""
@@ -77,7 +81,8 @@ def extract_embedded_json(s: str) -> dict:
     last_brace_index = s.rfind("}")
 
     if first_brace_index == -1 or last_brace_index == -1:
-        raise ValueError("No valid json found")
+        logger.warning("No valid json found, assuming answer is entire string")
+        return {"answer": s, "quotes": []}
 
     json_str = s[first_brace_index : last_brace_index + 1]
     try:

backend/tests/unit/danswer/llm/answering/stream_processing/test_quote_processing.py

@@ -324,8 +324,13 @@ def test_lengthy_prefixed_json_with_quotes() -> None:
     assert quotes[0] == "Document"
 
 
-def test_prefixed_json_with_quotes() -> None:
+def test_json_with_lengthy_prefix_and_quotes() -> None:
     tokens = [
+        "*** Based on the provided documents, there does not appear to be any information ",
+        "directly relevant to answering which documents are my favorite. ",
+        "The documents seem to be focused on describing the Danswer product ",
+        "and its features/use cases. Since I do not have personal preferences ",
+        "for documents, I will provide a general response:\n\n",
         "```",
         "json",
         "\n",