def normalize_ws(s: str) -> str:
   return re.sub(r"\s+", " ", s).strip()


RAW_DOCS = [
   {
       "doc_id": "trulens_core",
       "title": "TruLens core idea",
       "text": "TruLens is used to track and evaluate LLM applications. It can log app runs, compute feedback scores, and provide a dashboard to compare versions and investigate traces and results."
   },
   {
       "doc_id": "trulens_feedback",
       "title": "Feedback functions",
       "text": "TruLens feedback functions can score groundedness, context relevance, and answer relevance. They are configured by specifying which parts of an app record should be used as inputs."
   },
   {
       "doc_id": "trulens_rag",
       "title": "RAG workflow",
       "text": "A typical RAG system retrieves relevant chunks from a vector database and then generates an answer using those chunks as context. The quality depends on retrieval, prompt design, and generation behavior."
   },
   {
       "doc_id": "trulens_instrumentation",
       "title": "Instrumentation",
       "text": "Instrumentation adds tracing spans to your app functions (like retrieval and generation). This makes it possible to analyze which contexts were retrieved, latency, token usage, and connect feedback evaluations to specific steps."
   },
   {
       "doc_id": "vectorstores",
       "title": "Vector stores and embeddings",
       "text": "Vector stores index embeddings for text chunks, enabling semantic search. OpenAI embedding models can be used to embed chunks and queries, and Chroma can store them locally in memory for a notebook demo."
   },
   {
       "doc_id": "prompting",
       "title": "Prompting and citations",
       "text": "Prompting can encourage careful, citation-grounded answers. A stronger prompt can enforce: answer only from context, be explicit about uncertainty, and provide short citations that map to retrieved chunks."
   },
]


@dataclass
class Chunk:
   chunk_id: str
   doc_id: str
   title: str
   text: str
   meta: Dict[str, Any]


def chunk_docs(docs, chunk_size=350, overlap=80) -> List[Chunk]:
   chunks: List[Chunk] = []
   for d in docs:
       text = normalize_ws(d["text"])
       start = 0
       idx = 0
       while start < len(text):
           end = min(len(text), start + chunk_size)
           chunk_text = text[start:end]
           chunk_id = f'{d["doc_id"]}_c{idx}'
           chunks.append(
               Chunk(
                   chunk_id=chunk_id,
                   doc_id=d["doc_id"],
                   title=d["title"],
                   text=chunk_text,
                   meta={"doc_id": d["doc_id"], "title": d["title"], "chunk_index": idx},
               )
           )
           idx += 1
           start = end - overlap
           if start < 0:
               start = 0
           if end == len(text):
               break
   return chunks


CHUNKS = chunk_docs(RAW_DOCS)