Docs: Grammar action, satisfaction! First 20 pages. (#1817)

* update prompt
change model
fix chunking

* add sanity check and switch back to gpt 4

* fix lint
fix lint
fix grammar script
add offset limit functionailty for grammar checker

* first 20 pages

* Apply suggestions from code review

---------

Co-authored-by: Anton Burnashev <[email protected]>

docs/tools/fix_grammar_gpt.py

@@ -17,13 +17,15 @@
 
 # constants
 BASE_DIR = "../website/docs"
-GPT_MODEL = "gpt-3.5-turbo-0125"
-MAX_CHUNK_SIZE = 14000  # make sure that this is below the context window size of the model to not have cut off files
+GPT_MODEL = "gpt-4-turbo"
+MAX_CHUNK_SIZE = 4000  # make sure that this is below the context window size of the model to not have cut off files
 
 SYSTEM_PROMPT = """\
 You are a grammar checker. Every message you get will be a document that is to be grammarchecked and returned as such.
 You will not change the markdown syntax. You will only fix the grammar. You will not change the code snippets except for the comments therein.
 You will not modify the header section which is enclosed by two occurences of "---".
+Make sure all headings use the Sentence case.
+Never insert any codeblock start or end statements such as "```"
 Do not change the spelling or casing of these words: dlt, sdf, dbt
 """
 
@@ -51,6 +53,22 @@
         type=str,
     )
 
+    parser.add_argument(
+        "-o",
+        "--offset",
+        help="File count offset from where to start fixing",
+        default=0,
+        type=int,
+    )
+
+    parser.add_argument(
+        "-l",
+        "--limit",
+        help="File count limit, how many files to process",
+        default=100000,
+        type=int,
+    )
+
     # get args
     args = parser.parse_args()
 
@@ -63,14 +81,26 @@
 
     # run grammar check
     count = 0
+    processed = 0
     for file_path in markdown_files:
         count += 1
 
+        if count <= args.offset:
+            continue
+
+        if processed >= args.limit:
+            break
+
+        processed += 1
+
         fmt.note(f"Fixing grammar for file {file_path} ({count} of {len(markdown_files)})")
 
         with open(file_path, "r", encoding="utf-8") as f:
             doc = f.readlines()
 
+        with open(file_path, "r", encoding="utf-8") as f:
+            doc_length = len(f.read())
+
         def get_chunk_length(chunk: List[str]) -> int:
             count = 0
             for line in chunk:
@@ -80,8 +110,11 @@ def get_chunk_length(chunk: List[str]) -> int:
         # cut file into sections
         sections: List[List[str]] = []
         current_section: List[str] = []
+        is_in_code_block: bool = False
         for line in doc:
-            if line.startswith("#"):
+            if "```" in line:
+                is_in_code_block = not is_in_code_block
+            if line.startswith("#") and not is_in_code_block:
                 if current_section:
                     sections.append(current_section)
                 current_section = [line]
@@ -106,27 +139,42 @@ def get_chunk_length(chunk: List[str]) -> int:
         # sanity test, make sure we still have the full doc
         assert doc == functools.reduce(lambda a, b: a + b, chunks)
 
-        fmt.note(f"Created {len(chunks)} chunks")
+        # count chars in doc
+        fmt.note(f"Created {len(chunks)} chunks for {doc_length} chars")
 
-        fixed_chunks: List[List[str]] = []
+        fixed_chunks: List[str] = []
         for chunk in chunks:
             client = OpenAI()
+            in_string = "".join(chunk)
             response = client.chat.completions.create(
+                seed=123981298,
                 model=GPT_MODEL,
                 messages=[
                     {"role": "system", "content": SYSTEM_PROMPT},
-                    {"role": "user", "content": "".join(chunk)},
+                    {"role": "user", "content": in_string},
                 ],
                 temperature=0,
             )
-
-            fixed_chunks.append(response.choices[0].message.content)  # type: ignore
+            fixed_chunks.append(response.choices[0].message.content)
+
+        # here we check that no part of the doc was swallowed by gpt
+        fixed_doc_length = functools.reduce(
+            lambda count, chunk: count + len(chunk), fixed_chunks, 0
+        )
+        if fixed_doc_length / doc_length < 0.9:
+            fmt.error(
+                "Doc length reduced too much during processing, skipping saving, please check"
+                " manually"
+            )
+            continue
 
         with open(file_path, "w", encoding="utf-8") as f:
             for c in fixed_chunks:
-                f.writelines(c)
+                f.write(c)
+                f.write("\n")
+                f.write("\n")
 
     if count == 0:
         fmt.warning("No files selected for grammar check.")
     else:
-        fmt.note(f"Fixed grammar for {count} files.")
+        fmt.note(f"Fixed grammar for {processed} files.")

docs/website/docs/_book-onboarding-call.md

@@ -1 +1,2 @@
-<a href="https://calendar.app.google/EMZRS6YhM11zTGQw7">book a call</a> with a dltHub Solutions Engineer
+<a href="https://calendar.app.google/EMZRS6YhM11zTGQw7">Book a call</a> with a dltHub Solutions Engineer
+