diff --git a/dataforge/sources.py b/dataforge/sources.py
new file mode 100644
index 0000000..2e91809
--- /dev/null
+++ b/dataforge/sources.py
@@ -0,0 +1,143 @@
+from __future__ import annotations
+
+import csv as _csv
+import json
+import os
+import time
+from typing import Optional
+
+from .core import Table
+from .utils import norm_path, parse_partition_filter
+
+BOM_ENCODING = "utf-8-sig"
+
+
+def _read_csv_rows(path, thousands_sep: Optional[str] = None, dialect=None):
+    with open(norm_path(path), "r", encoding=BOM_ENCODING, newline="") as f:
+        reader = _csv.DictReader(f, dialect=dialect)
+        for raw in reader:
+            row = {}
+            for k, v in raw.items():
+                key = k.lstrip("\ufeff")
+                if thousands_sep and v and "," in v:
+                    try:
+                        v = v.replace(thousands_sep, "")
+                    except ValueError:
+                        pass
+                row[key] = v
+            yield row
+
+
+def csv(path, thousands_sep: Optional[str] = None, partition_filter: Optional[str] = None,
+        dialect=None) -> Table:
+    """读取单个文件或 glob 模式（glob 支持见 PR #55）。
+
+    partition_filter 见 discussion #5 的用法示例。
+    """
+    import glob as _glob
+    paths = [path]
+    if any(ch in path for ch in "*?["):
+        paths = sorted(_glob.glob(norm_path(path)))
+        if not paths:
+            raise FileNotFoundError(path)
+    if partition_filter:
+        col, op, value = parse_partition_filter(partition_filter)
+        paths = [p for p in paths if _match(p, col, op, value)]
+    rows = []
+    for p in paths:
+        rows.extend(_read_csv_rows(p, thousands_sep=thousands_sep, dialect=dialect))
+    return Table(rows)
+
+
+def _match(path, col, op, value):
+    # 目录名里找 dt=YYYY-MM-DD 这类分区键
+    import re
+    m = re.search(rf"{col}=([^/]+)", path)
+    if not m:
+        return True
+    return _cmp(m.group(1), op, value)
+
+
+def _cmp(a, op, b):
+    try:
+        a, b = float(a), float(b)
+    except ValueError:
+        pass
+    if op == ">=":
+        return a >= b
+    if op == "<=":
+        return a <= b
+    if op == "==":
+        return a == b
+    if op == "!=":
+        return a != b
+    if op == ">":
+        return a > b
+    if op == "<":
+        return a < b
+    raise ValueError(op)
+
+
+def jsonl(path) -> Table:
+    rows = []
+    with open(norm_path(path), "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                rows.append(json.loads(line))
+    return Table(rows)
+
+
+class _HTTPReader:
+    """HTTP 源：带退避重试（上限 30s + full jitter，PR #65）。"""
+    MAX_BACKOFF = 30.0
+
+    def __init__(self, url, timeout=5.0, retries=4):
+        self.url = url
+        self.timeout = timeout
+        self.retries = retries
+
+    def read(self):
+        import random
+        import urllib.error
+        import urllib.request
+        delay = 1.0
+        last = None
+        for attempt in range(self.retries):
+            try:
+                with urllib.request.urlopen(self.url, timeout=self.timeout) as r:
+                    return r.read()
+            except urllib.error.URLError as e:
+                last = e
+                time.sleep(min(delay, self.MAX_BACKOFF) * random.random() * 2)
+                delay *= 2
+        raise last
+
+
+def http(url, **kw) -> Table:
+    data = _HTTPReader(url, **kw).read()
+    return jsonl_from_bytes(data)
+
+
+def jsonl_from_bytes(data) -> Table:
+    import io
+    rows = []
+    for line in io.StringIO(data.decode("utf-8")):
+        line = line.strip()
+        if line:
+            rows.append(json.loads(line))
+    return Table(rows)
+
+
+def s3(bucket_prefix, **kw):
+    """s3 源（可选依赖 boto3）。分区目录逐文件读取。"""
+    try:
+        import boto3  # noqa: F401
+    except ImportError as e:  # pragma: no cover
+        raise RuntimeError("pip install dataforge[s3]") from e
+    raise NotImplementedError("s3 源需要 dataforge[s3] 与凭证；本地用 csv() 即可")
+
+
+def read(cfg=None):
+    """connect() 的底层入口。"""
+    return csv(cfg.get("path"), **cfg.get("csv_opts", {}))