metrics

evaluation/rag_metrics.py

@@ -17,6 +17,11 @@ MARKDOWN_URL_RE = re.compile(
     r"\[[^\]]*\]\((https?://[^)]+)\)"
 )
 
+WORD_RE = re.compile(
+    r"[^\W_]+",
+    re.UNICODE,
+)
+
 
 def normalize_text(
     value: str,
@@ -24,15 +29,175 @@ def normalize_text(
     value = unicodedata.normalize(
         "NFKC",
         value,
-    )
-
-    value = value.casefold()
+    ).casefold()
 
     return " ".join(
         value.split()
     )
 
 
+def normalize_match_token(
+    value: str,
+) -> str:
+    value = unicodedata.normalize(
+        "NFKD",
+        value,
+    )
+
+    value = "".join(
+        character
+        for character in value
+        if not unicodedata.combining(
+            character
+        )
+    )
+
+    return value.casefold()
+
+
+def match_tokens(
+    value: str,
+) -> list[str]:
+    normalized = (
+        unicodedata.normalize(
+            "NFKC",
+            value,
+        )
+    )
+
+    return [
+        normalize_match_token(
+            token
+        )
+        for token
+        in WORD_RE.findall(
+            normalized
+        )
+    ]
+
+
+def morphology_token_matches(
+    expected: str,
+    actual: str,
+) -> bool:
+    if expected == actual:
+        return True
+
+    if (
+        expected.isdigit()
+        or actual.isdigit()
+    ):
+        return False
+
+    shorter = min(
+        len(expected),
+        len(actual),
+    )
+
+    if shorter <= 2:
+        return False
+
+    required_prefix = (
+        3
+        if shorter <= 4
+        else 4
+        if shorter <= 6
+        else 5
+    )
+
+    return (
+        expected[
+            :required_prefix
+        ]
+        == actual[
+            :required_prefix
+        ]
+    )
+
+
+def expected_phrase_matches(
+    expected: str,
+    answer: str,
+) -> bool:
+    normalized_expected = (
+        normalize_text(
+            expected
+        )
+    )
+
+    normalized_answer = (
+        normalize_text(
+            answer
+        )
+    )
+
+    if not normalized_expected:
+        return True
+
+    expected_tokens = (
+        match_tokens(
+            expected
+        )
+    )
+
+    answer_tokens = (
+        match_tokens(
+            answer
+        )
+    )
+
+    has_numeric_token = any(
+        token.isdigit()
+        for token in expected_tokens
+    )
+
+    if (
+        not has_numeric_token
+        and normalized_expected
+        in normalized_answer
+    ):
+        return True
+
+    if (
+        not expected_tokens
+        or len(answer_tokens)
+        < len(expected_tokens)
+    ):
+        return False
+
+    window_size = len(
+        expected_tokens
+    )
+
+    for start in range(
+        len(answer_tokens)
+        - window_size
+        + 1
+    ):
+        window = answer_tokens[
+            start:
+            start + window_size
+        ]
+
+        if all(
+            morphology_token_matches(
+                expected_token,
+                actual_token,
+            )
+            for (
+                expected_token,
+                actual_token,
+            )
+            in zip(
+                expected_tokens,
+                window,
+            )
+        ):
+            return True
+
+    return False
+
+
 def normalize_url(
     value: str,
 ) -> str:
@@ -79,25 +244,15 @@ def evaluate_answer(
     ):
         expected_contains = []
 
-    answer_matches: list[
-        bool
-    ] = []
+    answer_matches = [
+        expected_phrase_matches(
+            str(expected),
+            answer,
+        )
+        for expected
+        in expected_contains
+    ]
 
-    for expected in expected_contains:
-        expected_text = (
-            normalize_text(
-                str(
-                    expected
-                )
-            )
-        )
-
-        answer_matches.append(
-            expected_text
-            in normalized_answer
-        )
-
-    if answer_matches:
     answer_contains_score = (
         sum(
             answer_matches
@@ -105,11 +260,10 @@ def evaluate_answer(
         / len(
             answer_matches
         )
+        if answer_matches
+        else 1.0
     )
 
-    else:
-        answer_contains_score = 1.0
-
     expected_urls = (
         question.get(
             "expected_source_urls",
@@ -125,26 +279,18 @@ def evaluate_answer(
 
     normalized_expected_urls = [
         normalize_url(
-            str(
-                url
-            )
+            str(url)
         )
         for url in expected_urls
     ]
 
-    source_matches: list[
-        bool
-    ] = []
-
-    for expected_url in (
-        normalized_expected_urls
-    ):
-        source_matches.append(
+    source_matches = [
         expected_url
         in answer
-        )
+        for expected_url
+        in normalized_expected_urls
+    ]
 
-    if source_matches:
     source_url_score = (
         sum(
             source_matches
@@ -152,11 +298,10 @@ def evaluate_answer(
         / len(
             source_matches
         )
+        if source_matches
+        else 1.0
     )
 
-    else:
-        source_url_score = 1.0
-
     should_answer = bool(
         question.get(
             "should_answer",
@@ -164,14 +309,10 @@ def evaluate_answer(
         )
     )
 
-    normalized_no_answer = (
+    returned_no_answer = (
         normalize_text(
             NO_ANSWER_TEXT
         )
-    )
-
-    returned_no_answer = (
-        normalized_no_answer
         in normalized_answer
     )
 
@@ -539,7 +680,8 @@ def aggregate_by_field(
         for (
             group_name,
             group_rows,
-        ) in sorted(
+        )
+        in sorted(
             grouped.items()
         )
     }