ClareCourseWare / api /rag_engine.py
claudqunwang's picture
Add Clare product UI: run_web.sh, README, exclude hf_space from push
c8c6034
# api/rag_engine.py
"""
RAG engine with vector database support:
- build_rag_chunks_from_file(path, doc_type) -> List[chunk] (with embeddings)
- retrieve_relevant_chunks(query, chunks, ...) -> (context_text, used_chunks)
- Uses FAISS vector similarity + token overlap rerank
Chunk format (enhanced):
{
"text": str,
"source_file": str,
"section": str,
"doc_type": str,
"embedding": Optional[List[float]] # NEW: OpenAI embedding vector
}
PDF parsing:
- Priority: unstructured.io (better quality)
- Fallback: pypdf (if unstructured fails)
"""
import os
import re
import math
from typing import Dict, List, Tuple, Optional, Any
# Legacy parsers (fallback)
from pypdf import PdfReader
from docx import Document
from pptx import Presentation
# Embedding & vector DB
from .config import client, EMBEDDING_MODEL
from .clare_core import cosine_similarity
# ============================
# Optional: Better PDF parsing (unstructured.io)
# ============================
def _safe_import_unstructured():
try:
from unstructured.partition.auto import partition
return partition
except Exception:
try:
# Fallback to older API
from unstructured.partition.pdf import partition_pdf
return partition_pdf
except Exception:
return None
# ============================
# Optional: FAISS vector database
# ============================
def _safe_import_faiss():
try:
import faiss # type: ignore
return faiss
except Exception:
return None
# ============================
# Token helpers (optional tiktoken)
# ============================
def _safe_import_tiktoken():
try:
import tiktoken # type: ignore
return tiktoken
except Exception:
return None
def _approx_tokens(text: str) -> int:
if not text:
return 0
return max(1, int(len(text) / 4))
def _count_text_tokens(text: str, model: str = "") -> int:
tk = _safe_import_tiktoken()
if tk is None:
return _approx_tokens(text)
try:
enc = tk.encoding_for_model(model) if model else tk.get_encoding("cl100k_base")
except Exception:
enc = tk.get_encoding("cl100k_base")
return len(enc.encode(text or ""))
def _truncate_to_tokens(text: str, max_tokens: int, model: str = "") -> str:
"""Deterministic truncation. Uses tiktoken if available; otherwise approximates by char ratio."""
if not text:
return text
tk = _safe_import_tiktoken()
if tk is None:
total = _approx_tokens(text)
if total <= max_tokens:
return text
ratio = max_tokens / max(1, total)
cut = max(50, min(len(text), int(len(text) * ratio)))
s = text[:cut]
while _approx_tokens(s) > max_tokens and len(s) > 50:
s = s[: int(len(s) * 0.9)]
return s
try:
enc = tk.encoding_for_model(model) if model else tk.get_encoding("cl100k_base")
except Exception:
enc = tk.get_encoding("cl100k_base")
ids = enc.encode(text or "")
if len(ids) <= max_tokens:
return text
return enc.decode(ids[:max_tokens])
# ============================
# RAG hard limits
# ============================
RAG_TOPK_LIMIT = 4
RAG_CHUNK_TOKEN_LIMIT = 500
RAG_CONTEXT_TOKEN_LIMIT = 2000 # 4 * 500
# Embedding dimension for text-embedding-3-small
EMBEDDING_DIM = 1536
# ----------------------------
# Helpers
# ----------------------------
def _clean_text(s: str) -> str:
s = (s or "").replace("\r", "\n")
s = re.sub(r"\n{3,}", "\n\n", s)
return s.strip()
def _split_into_chunks(text: str, max_chars: int = 1400) -> List[str]:
"""
Simple deterministic chunker:
- split by blank lines
- then pack into <= max_chars
"""
text = _clean_text(text)
if not text:
return []
paras = [p.strip() for p in text.split("\n\n") if p.strip()]
chunks: List[str] = []
buf = ""
for p in paras:
if not buf:
buf = p
continue
if len(buf) + 2 + len(p) <= max_chars:
buf = buf + "\n\n" + p
else:
chunks.append(buf)
buf = p
if buf:
chunks.append(buf)
return chunks
def _file_label(path: str) -> str:
return os.path.basename(path) if path else "uploaded_file"
def _basename(x: str) -> str:
try:
return os.path.basename(x or "")
except Exception:
return x or ""
# ----------------------------
# Embedding generation
# ----------------------------
def get_chunk_embedding(text: str) -> Optional[List[float]]:
"""Generate embedding for a chunk using OpenAI text-embedding-3-small."""
if not text or not text.strip():
return None
try:
resp = client.embeddings.create(
model=EMBEDDING_MODEL,
input=[text.strip()],
)
return resp.data[0].embedding
except Exception as e:
print(f"[rag_engine] embedding error: {repr(e)}")
return None
def get_chunk_embeddings_batch(texts: List[str], batch_size: int = 100) -> List[Optional[List[float]]]:
"""
Generate embeddings for multiple chunks in batches (more efficient than individual calls).
OpenAI API supports up to 2048 inputs per request, but we use smaller batches for reliability.
"""
if not texts:
return []
results: List[Optional[List[float]]] = []
for i in range(0, len(texts), batch_size):
batch = [t.strip() for t in texts[i:i + batch_size] if t and t.strip()]
if not batch:
results.extend([None] * (i + batch_size - len(results)))
continue
try:
resp = client.embeddings.create(
model=EMBEDDING_MODEL,
input=batch,
)
batch_results = [item.embedding for item in resp.data]
results.extend(batch_results)
except Exception as e:
print(f"[rag_engine] batch embedding error: {repr(e)}")
results.extend([None] * len(batch))
return results
# ----------------------------
# Enhanced PDF parsing (unstructured.io + fallback)
# ----------------------------
def _parse_pdf_to_text(path: str) -> List[Tuple[str, str]]:
"""
Returns list of (section_label, text)
Priority: unstructured.io (better quality)
Fallback: pypdf
"""
partition_func = _safe_import_unstructured()
# Try unstructured.io first
if partition_func is not None:
try:
# Try new API first (partition function)
if hasattr(partition_func, '__name__') and partition_func.__name__ == 'partition':
elements = partition_func(filename=path)
else:
# Old API (partition_pdf)
elements = partition_func(filename=path)
text_parts: List[str] = []
for elem in elements:
if hasattr(elem, "text") and elem.text:
text_parts.append(str(elem.text).strip())
if text_parts:
full_text = "\n\n".join(text_parts)
full_text = _clean_text(full_text)
if full_text:
return [("pdf_unstructured", full_text)]
except Exception as e:
print(f"[rag_engine] unstructured.io parse failed, fallback to pypdf: {repr(e)}")
# Fallback: pypdf
try:
reader = PdfReader(path)
out: List[Tuple[str, str]] = []
for i, page in enumerate(reader.pages):
t = page.extract_text() or ""
t = _clean_text(t)
if t:
out.append((f"p{i+1}", t))
return out
except Exception as e:
print(f"[rag_engine] pypdf parse error: {repr(e)}")
return []
def _parse_docx_to_text(path: str) -> List[Tuple[str, str]]:
doc = Document(path)
paras = [p.text.strip() for p in doc.paragraphs if p.text and p.text.strip()]
if not paras:
return []
full = "\n\n".join(paras)
return [("docx", _clean_text(full))]
def _parse_pptx_to_text(path: str) -> List[Tuple[str, str]]:
prs = Presentation(path)
out: List[Tuple[str, str]] = []
for idx, slide in enumerate(prs.slides, start=1):
lines: List[str] = []
for shape in slide.shapes:
if hasattr(shape, "text") and shape.text:
txt = shape.text.strip()
if txt:
lines.append(txt)
if lines:
out.append((f"slide{idx}", _clean_text("\n".join(lines))))
return out
# ----------------------------
# Vector database (FAISS) wrapper
# ----------------------------
class VectorStore:
"""Simple in-memory vector store using FAISS (or fallback to list-based cosine similarity)."""
def __init__(self):
self.faiss = _safe_import_faiss()
self.index = None
self.chunks: List[Dict] = []
self.use_faiss = False
def build_index(self, chunks: List[Dict]):
"""Build FAISS index from chunks with embeddings."""
self.chunks = chunks or []
if not self.chunks:
return
# Filter chunks that have embeddings
chunks_with_emb = [c for c in self.chunks if c.get("embedding") is not None]
if not chunks_with_emb:
print("[rag_engine] No chunks with embeddings, using token-based retrieval")
return
if self.faiss is None:
print("[rag_engine] FAISS not available, using list-based cosine similarity")
return
try:
dim = len(chunks_with_emb[0]["embedding"])
# Use L2 (Euclidean) index for FAISS
self.index = self.faiss.IndexFlatL2(dim)
embeddings = [c["embedding"] for c in chunks_with_emb]
import numpy as np
vectors = np.array(embeddings, dtype=np.float32)
self.index.add(vectors)
self.use_faiss = True
print(f"[rag_engine] Built FAISS index with {len(chunks_with_emb)} vectors")
except Exception as e:
print(f"[rag_engine] FAISS index build failed: {repr(e)}, using list-based")
self.use_faiss = False
def search(self, query_embedding: List[float], k: int) -> List[Tuple[float, Dict]]:
"""
Search top-k chunks by vector similarity.
Returns: List[(similarity_score, chunk_dict)]
"""
if not query_embedding or not self.chunks:
return []
chunks_with_emb = [c for c in self.chunks if c.get("embedding") is not None]
if not chunks_with_emb:
return []
if self.use_faiss and self.index is not None:
try:
import numpy as np
query_vec = np.array([query_embedding], dtype=np.float32)
distances, indices = self.index.search(query_vec, min(k, len(chunks_with_emb)))
results: List[Tuple[float, Dict]] = []
for dist, idx in zip(distances[0], indices[0]):
if idx < len(chunks_with_emb):
# Convert L2 distance to similarity (1 / (1 + distance))
similarity = 1.0 / (1.0 + float(dist))
results.append((similarity, chunks_with_emb[idx]))
return results
except Exception as e:
print(f"[rag_engine] FAISS search error: {repr(e)}, fallback to list-based")
# Fallback: list-based cosine similarity
results: List[Tuple[float, Dict]] = []
for chunk in chunks_with_emb:
emb = chunk.get("embedding")
if emb:
sim = cosine_similarity(query_embedding, emb)
results.append((sim, chunk))
results.sort(key=lambda x: x[0], reverse=True)
return results[:k]
# ----------------------------
# Public API
# ----------------------------
def build_rag_chunks_from_file(path: str, doc_type: str, generate_embeddings: bool = True) -> List[Dict]:
"""
Build RAG chunks from a local file path.
Supports: .pdf / .docx / .pptx / .txt
Args:
path: File path
doc_type: Document type
generate_embeddings: If True, generate embeddings for each chunk (default: True)
Returns:
List of chunk dicts with optional "embedding" field
"""
if not path or not os.path.exists(path):
return []
ext = os.path.splitext(path)[1].lower()
source_file = _file_label(path)
sections: List[Tuple[str, str]] = []
try:
if ext == ".pdf":
sections = _parse_pdf_to_text(path)
elif ext == ".docx":
sections = _parse_docx_to_text(path)
elif ext == ".pptx":
sections = _parse_pptx_to_text(path)
elif ext in [".txt", ".md"]:
with open(path, "r", encoding="utf-8", errors="ignore") as f:
sections = [("text", _clean_text(f.read()))]
else:
print(f"[rag_engine] unsupported file type: {ext}")
return []
except Exception as e:
print(f"[rag_engine] parse error for {source_file}: {repr(e)}")
return []
chunks: List[Dict] = []
chunk_texts: List[str] = []
# First, build all chunks without embeddings
for section, text in sections:
for j, piece in enumerate(_split_into_chunks(text), start=1):
chunk: Dict[str, Any] = {
"text": piece,
"source_file": source_file,
"section": f"{section}#{j}",
"doc_type": doc_type,
}
chunks.append(chunk)
if generate_embeddings:
chunk_texts.append(piece)
# Generate embeddings in batch (much faster than individual calls)
if generate_embeddings and chunk_texts:
embeddings = get_chunk_embeddings_batch(chunk_texts, batch_size=100)
for chunk, embedding in zip(chunks, embeddings):
if embedding:
chunk["embedding"] = embedding
return chunks
def retrieve_relevant_chunks(
query: str,
chunks: List[Dict],
k: int = RAG_TOPK_LIMIT,
max_context_chars: int = 600,
min_score: int = 6,
chunk_token_limit: int = RAG_CHUNK_TOKEN_LIMIT,
max_context_tokens: int = RAG_CONTEXT_TOKEN_LIMIT,
model_for_tokenizer: str = "",
allowed_source_files: Optional[List[str]] = None,
allowed_doc_types: Optional[List[str]] = None,
use_vector_search: bool = True, # NEW: enable/disable vector search
vector_similarity_threshold: float = 0.7, # Minimum cosine similarity for vector results
) -> Tuple[str, List[Dict]]:
"""
Enhanced retrieval with vector similarity + token overlap rerank.
Strategy:
1. If use_vector_search=True and chunks have embeddings:
- Generate query embedding
- Use FAISS/list-based vector similarity to get candidate chunks
- Rerank by token overlap
2. Else: fallback to token-based retrieval (backward compatible)
Args:
use_vector_search: Enable vector similarity search (default: True)
vector_similarity_threshold: Minimum cosine similarity for vector results (default: 0.7)
"""
query = _clean_text(query)
if not query or not chunks:
return "", []
# ----------------------------
# Apply scoping BEFORE scoring
# ----------------------------
filtered = chunks or []
if allowed_source_files:
allow_files = {_basename(str(x)).strip() for x in allowed_source_files if str(x).strip()}
if allow_files:
filtered = [
c
for c in filtered
if _basename(str(c.get("source_file", ""))).strip() in allow_files
]
if allowed_doc_types:
allow_dt = {str(x).strip() for x in allowed_doc_types if str(x).strip()}
if allow_dt:
filtered = [c for c in filtered if str(c.get("doc_type", "")).strip() in allow_dt]
if not filtered:
return "", []
# Short query gate
q_tokens_list = re.findall(r"[a-zA-Z0-9]+", query.lower())
if (len(q_tokens_list) < 3) and (len(query) < 20):
return "", []
q_tokens = set(q_tokens_list)
if not q_tokens:
return "", []
# ----------------------------
# Vector search path (if enabled and embeddings available)
# ----------------------------
chunks_with_emb = [c for c in filtered if c.get("embedding") is not None]
if use_vector_search and chunks_with_emb:
try:
query_emb = get_chunk_embedding(query)
if query_emb:
# Build vector store and search
store = VectorStore()
store.build_index(chunks_with_emb)
vector_results = store.search(query_emb, k=k * 2) # Get 2x candidates for rerank
# Filter by similarity threshold
candidates: List[Tuple[float, Dict]] = []
for sim_score, chunk in vector_results:
if float(sim_score) >= vector_similarity_threshold:
candidates.append((float(sim_score), chunk))
if candidates:
# Rerank by token overlap
scored: List[Tuple[float, Dict]] = []
for sim_score, c in candidates:
text = (c.get("text") or "")
if not text:
continue
t_tokens = set(re.findall(r"[a-zA-Z0-9]+", text.lower()))
token_score = len(q_tokens.intersection(t_tokens))
token_ratio = min(1.0, float(token_score) / max(1, len(q_tokens)))
# Combined score: 70% vector similarity + 30% token overlap (normalized)
combined_score = 0.7 * float(sim_score) + 0.3 * token_ratio
c2 = dict(c)
c2["_rag_vector_sim"] = float(sim_score)
c2["_rag_token_overlap"] = int(token_score)
c2["_rag_token_overlap_ratio"] = float(token_ratio)
c2["_rag_score"] = float(combined_score)
scored.append((combined_score, c2))
scored.sort(key=lambda x: x[0], reverse=True)
top = [c for _, c in scored[:k]]
else:
# Vector search found nothing above threshold, fallback to token
top = []
else:
top = []
except Exception as e:
print(f"[rag_engine] vector search error: {repr(e)}, fallback to token-based")
top = []
else:
top = []
# If vector search returns unrelated chunks (e.g. zero token overlap), treat as no-hit and fallback.
if top:
doc_hint_tokens = {
"module", "week", "lab", "assignment", "syllabus", "lecture", "slide", "ppt", "pdf", "docx",
"课程", "模块", "周", "实验", "作业", "讲义", "课件", "大纲", "论文",
}
looks_like_course_query = any(t in query.lower() for t in doc_hint_tokens)
best_overlap = max(int(c.get("_rag_token_overlap", 0)) for c in top)
best_score = max(float(c.get("_rag_score", 0.0)) for c in top)
if (not looks_like_course_query and best_overlap <= 0) or best_score < 0.35:
top = []
# ----------------------------
# Fallback: token-based retrieval (if vector search failed or disabled)
# ----------------------------
if not top:
scored: List[Tuple[int, Dict]] = []
for c in filtered:
text = (c.get("text") or "")
if not text:
continue
t_tokens = set(re.findall(r"[a-zA-Z0-9]+", text.lower()))
score = len(q_tokens.intersection(t_tokens))
if score >= min_score:
scored.append((score, c))
if not scored:
return "", []
scored.sort(key=lambda x: x[0], reverse=True)
k_actual = min(int(k or RAG_TOPK_LIMIT), RAG_TOPK_LIMIT)
top = [c for _, c in scored[:k_actual]]
if not top:
return "", []
# ----------------------------
# Truncate and format context
# ----------------------------
used: List[Dict] = []
truncated_texts: List[str] = []
total_tokens = 0
for c in top:
raw = c.get("text") or ""
if not raw:
continue
t = _truncate_to_tokens(raw, max_tokens=chunk_token_limit, model=model_for_tokenizer)
t_tokens = _count_text_tokens(t, model=model_for_tokenizer)
if total_tokens + t_tokens > max_context_tokens:
remaining = max_context_tokens - total_tokens
if remaining <= 0:
break
t = _truncate_to_tokens(t, max_tokens=remaining, model=model_for_tokenizer)
t_tokens = _count_text_tokens(t, model=model_for_tokenizer)
if max_context_chars and max_context_chars > 0:
current_chars = sum(len(x) for x in truncated_texts)
if current_chars + len(t) > max_context_chars:
t = t[: max(0, max_context_chars - current_chars)]
t = _clean_text(t)
if not t:
continue
truncated_texts.append(t)
used.append(c)
total_tokens += t_tokens
if total_tokens >= max_context_tokens:
break
if not truncated_texts:
return "", []
context = "\n\n---\n\n".join(truncated_texts)
return context, used