fix: prompt-injection delimiter escape, legacy config coercion, gather resilience, true cross-thread concurrency bound

Addresses items 4-8 from the max-effort review of PR #272 (VectifyAI/PageIndex#272).

- agent.py: wrap_with_doc_context() strips '<'/'>' from doc_name/doc_description
  (untrusted: unsanitized filename / LLM-generated from document content) before
  inserting them into the <docs>...</docs> block, so embedded content can never
  form a literal </docs> that closes the delimiter early and escapes the
  untrusted-data boundary SCOPED_SYSTEM_PROMPT relies on. Deterministic
  per-field transform, doesn't touch the (cacheable) static system prompt.

- ConfigLoader.load() (legacy 0.2.x compat) now routes merged overrides through
  IndexConfig before returning, so a legacy 'no' string gets pydantic's bool
  coercion instead of surviving as a truthy non-empty string — page_index_main's
  bare `if opt.if_add_node_summary:` checks (changed from `== 'yes'` elsewhere
  in this PR) were silently inverting caller intent and firing unwanted billed
  LLM calls.

- verify_toc, process_large_node_recursively, tree_parser,
  generate_summaries_for_structure, generate_summaries_for_structure_md: added
  return_exceptions=True to their asyncio.gather calls (llm_completion/
  llm_acompletion raise RuntimeError on retry exhaustion, added earlier in this
  PR), each with a degrade path matching the pattern already used by sibling
  hardened gathers in the same files. One transient LLM failure no longer
  aborts the whole document's indexing.

- _llm_semaphore is now a true process-wide ceiling (threading.Semaphore,
  shared across every thread/event loop) instead of one asyncio.Semaphore per
  event loop -- concurrently indexing N documents on N threads no longer
  multiplies the effective cap by N. A max_concurrency_scope() override is
  layered as a second, nested, per-loop restriction that can only tighten the
  effective cap within the ceiling, never widen past it.

- set_llm_params() mutated a bare process-wide dict with no per-call isolation,
  unlike max_concurrency which already had ContextVar scoping. Added
  llm_params_scope() (mirrors max_concurrency_scope) + IndexConfig.llm_params,
  wired into build_index() the same way max_concurrency already was, so
  concurrent indexing jobs with different llm kwargs don't leak into each
  other.

Adds regression tests for all five. Full suite: 221 passed, 2 skipped (one
pre-existing, unrelated flaky cloud-streaming test intermittently fails on
rerun; confirmed independent of this change).

Claude-Session: https://claude.ai/code/session_01Kx5DgKbhK1N8autqXH8SmS

pageindex/agent.py

@@ -46,20 +46,29 @@ Answer based only on tool output. Be concise.
 """
 
 
+def _defang_delimiters(text: str) -> str:
+    """Strip '<'/'>' so untrusted text can never form a literal <docs>/</docs>
+    (or any other tag-shaped string) that would prematurely close the
+    wrap_with_doc_context() delimiter and escape the untrusted-data boundary."""
+    return text.replace("<", "").replace(">", "")
+
+
 def wrap_with_doc_context(docs: list[dict], question: str) -> str:
     """Prepend a doc-context block to the user question for scoped queries.
 
     Document fields (especially doc_description, which is LLM-generated at
     index time) are untrusted text that may contain adversarial instructions.
     We wrap them in a <docs>...</docs> delimiter and tell the agent in the
-    system prompt to treat the block as data only.
+    system prompt to treat the block as data only. '<'/'>' are stripped from
+    the untrusted fields first so embedded content can never form a literal
+    </docs> (or any other tag) that closes the delimiter early.
     """
     lines = []
     for d in docs:
-        line = f"- {d['doc_id']}: {d.get('doc_name', '')}"
+        line = f"- {d['doc_id']}: {_defang_delimiters(d.get('doc_name', ''))}"
         desc = d.get("doc_description") or ""
         if desc:
-            line += f" — {desc}"
+            line += f" — {_defang_delimiters(desc)}"
         lines.append(line)
     label = "document" if len(docs) == 1 else "documents"
     return (

pageindex/config.py

@@ -29,6 +29,11 @@ class IndexConfig(BaseModel):
     # default (get_max_concurrency(), overridable via PAGEINDEX_MAX_CONCURRENCY).
     # An explicit value here wins for this client.
     max_concurrency: int | None = None
+    # Per-call litellm completion kwargs for this client's indexing calls only
+    # (e.g. {"temperature": 1}). None = use the process-wide defaults
+    # (get_llm_params(), overridable via set_llm_params()). Scoped via
+    # llm_params_scope so it doesn't leak into other concurrent indexing calls.
+    llm_params: dict | None = None
 
     @field_validator("max_concurrency", mode="before")
     @classmethod
@@ -57,6 +62,16 @@ def _env_drop_params_default() -> bool:
 # PAGEINDEX_DROP_PARAMS env shortcut.
 _LLM_PARAMS: dict = {"temperature": 0, "drop_params": _env_drop_params_default()}
 
+# Per-call override, isolated per thread / async context — mirrors
+# _MAX_CONCURRENCY_OVERRIDE below. Without this, set_llm_params() is the only
+# way to change llm params and it mutates the process-wide dict directly, so
+# concurrently indexing two documents with different llm_params_scope() would
+# otherwise leak one caller's settings (e.g. temperature) into the other's
+# in-flight calls. None = no override -> fall back to the process-wide _LLM_PARAMS.
+_LLM_PARAMS_OVERRIDE: ContextVar[dict | None] = ContextVar(
+    "pageindex_llm_params_override", default=None
+)
+
 # Structural kwargs PageIndex always supplies itself — not overridable here.
 _RESERVED_LLM_PARAMS = ("model", "messages")
 
@@ -121,6 +136,16 @@ def get_max_concurrency() -> int:
     return override if override is not None else _MAX_CONCURRENCY
 
 
+def _process_wide_max_concurrency() -> int:
+    """The process-wide default cap, ignoring any active max_concurrency_scope
+    override. This is the TRUE ceiling shared across every thread/event loop in
+    the process (see index/utils.py's _llm_semaphore) — a per-call override may
+    only narrow the effective cap within that ceiling, never widen it, so the
+    ceiling itself must not vary with a context-local override.
+    """
+    return _MAX_CONCURRENCY
+
+
 def set_max_concurrency(value: int) -> None:
     """Set the process-wide default cap on concurrent in-flight LLM calls."""
     global _MAX_CONCURRENCY
@@ -147,19 +172,53 @@ def max_concurrency_scope(value: int | None):
 
 
 def get_llm_params() -> dict:
-    """Return a copy of the per-call kwargs PageIndex passes to litellm."""
-    return dict(_LLM_PARAMS)
+    """Return a copy of the effective per-call kwargs PageIndex passes to litellm.
+
+    A per-index override (llm_params_scope) is merged over the process-wide
+    defaults for the current context; otherwise just the process-wide defaults
+    apply.
+    """
+    params = dict(_LLM_PARAMS)
+    override = _LLM_PARAMS_OVERRIDE.get()
+    if override:
+        params.update(override)
+    return params
 
 
 def set_llm_params(**kwargs) -> None:
-    """Override or extend the litellm completion kwargs PageIndex sends per call.
+    """Override or extend the process-wide default litellm completion kwargs.
 
     e.g. ``set_llm_params(drop_params=False, temperature=1, num_retries=5)``.
-    Applied per call; never writes litellm's global state, so it can't leak into
-    other litellm users in the same process. ``model`` / ``messages`` are
-    reserved (PageIndex supplies them) and rejected.
+    Never writes litellm's global state, so it can't leak into other litellm
+    users in the same process — but it DOES mutate PageIndex's own process-wide
+    default, so it affects every concurrent caller in this process. For a
+    one-off override scoped to a single indexing call, use ``llm_params_scope``
+    instead. ``model`` / ``messages`` are reserved (PageIndex supplies them) and
+    rejected.
     """
     reserved = [k for k in kwargs if k in _RESERVED_LLM_PARAMS]
     if reserved:
         raise ValueError(f"cannot override reserved litellm kwargs: {reserved}")
     _LLM_PARAMS.update(kwargs)
+
+
+@contextmanager
+def llm_params_scope(overrides: dict | None):
+    """Scope a per-index override of the litellm completion kwargs to the
+    current context.
+
+    ``overrides=None`` (or ``{}``) means "no override" (fall back to the
+    process-wide defaults). Isolated per thread / async context and reset on
+    exit, so concurrent indexing doesn't leak one call's kwargs into another's
+    and a one-off override never becomes the sticky new process default —
+    mirrors ``max_concurrency_scope``.
+    """
+    if overrides:
+        reserved = [k for k in overrides if k in _RESERVED_LLM_PARAMS]
+        if reserved:
+            raise ValueError(f"cannot override reserved litellm kwargs: {reserved}")
+    token = _LLM_PARAMS_OVERRIDE.set(overrides or None)
+    try:
+        yield
+    finally:
+        _LLM_PARAMS_OVERRIDE.reset(token)

pageindex/index/page_index.py

@@ -929,13 +929,23 @@ async def verify_toc(page_list, list_result, start_index=1, N=None, model=None):
             item_with_index['list_index'] = idx  # Add the original index in list_result
             indexed_sample_list.append(item_with_index)
 
-    # Run checks concurrently
+    # Run checks concurrently. return_exceptions=True: a transient LLM failure
+    # on one sampled item must degrade that item to 'no' (same as an
+    # unavailable physical_index above), not abort verification for the
+    # whole document.
     tasks = [
         check_title_appearance(item, page_list, start_index, model)
         for item in indexed_sample_list
     ]
-    results = await asyncio.gather(*tasks)
-    
+    raw_results = await asyncio.gather(*tasks, return_exceptions=True)
+    results = []
+    for item, result in zip(indexed_sample_list, raw_results):
+        if isinstance(result, Exception):
+            results.append({'list_index': item.get('list_index'), 'answer': 'no',
+                            'title': item.get('title'), 'page_number': item.get('physical_index')})
+        else:
+            results.append(result)
+
     # Process results
     correct_count = 0
     incorrect_results = []
@@ -1022,8 +1032,14 @@ async def process_large_node_recursively(node, page_list, opt=None, logger=None)
             process_large_node_recursively(child_node, page_list, opt, logger=logger)
             for child_node in node['nodes']
         ]
-        await asyncio.gather(*tasks)
-    
+        # return_exceptions=True: one child subtree failing to expand further
+        # must not abort the whole document — it's left as a leaf at its
+        # current boundaries instead.
+        results = await asyncio.gather(*tasks, return_exceptions=True)
+        for child_node, result in zip(node['nodes'], results):
+            if isinstance(result, Exception) and logger:
+                logger.error(f"Failed to expand node '{child_node.get('title')}': {result}")
+
     return node
 
 async def tree_parser(page_list, opt, doc=None, logger=None):
@@ -1058,8 +1074,13 @@ async def tree_parser(page_list, opt, doc=None, logger=None):
         process_large_node_recursively(node, page_list, opt, logger=logger)
         for node in toc_tree
     ]
-    await asyncio.gather(*tasks)
-    
+    # return_exceptions=True: one top-level node failing to expand further
+    # must not abort indexing the whole document.
+    results = await asyncio.gather(*tasks, return_exceptions=True)
+    for node, result in zip(toc_tree, results):
+        if isinstance(result, Exception) and logger:
+            logger.error(f"Failed to expand node '{node.get('title')}': {result}")
+
     return toc_tree
 
 

pageindex/index/page_index_md.py

@@ -16,8 +16,14 @@ async def get_node_summary(node, summary_token_threshold=200, model=None):
 async def generate_summaries_for_structure_md(structure, summary_token_threshold, model=None):
     nodes = structure_to_list(structure)
     tasks = [get_node_summary(node, summary_token_threshold=summary_token_threshold, model=model) for node in nodes]
-    summaries = await asyncio.gather(*tasks)
-    
+    # return_exceptions=True: one node's summary failing must not abort
+    # summarization for the whole document — fall back to its raw text.
+    raw_summaries = await asyncio.gather(*tasks, return_exceptions=True)
+    summaries = [
+        node.get('text', '') if isinstance(s, Exception) else s
+        for node, s in zip(nodes, raw_summaries)
+    ]
+
     for node, summary in zip(nodes, summaries):
         if not node.get('nodes'):
             node['summary'] = summary

pageindex/index/pipeline.py

@@ -68,15 +68,16 @@ def build_index(parsed: ParsedDocument, model: str = None, opt=None) -> dict:
     from .utils import (write_node_id, add_node_text, remove_structure_text,
                         generate_summaries_for_structure, generate_doc_description,
                         create_clean_structure_for_description)
-    from ..config import IndexConfig, max_concurrency_scope
+    from ..config import IndexConfig, max_concurrency_scope, llm_params_scope
 
     if opt is None:
         opt = IndexConfig(model=model) if model else IndexConfig()
 
-    # Scope the per-index concurrency cap to THIS call only (per thread/async
-    # context), so concurrent indexing of other documents isn't affected and a
-    # one-off value never sticks as the process default.
-    with max_concurrency_scope(getattr(opt, "max_concurrency", None)):
+    # Scope the per-index concurrency cap AND llm kwargs to THIS call only (per
+    # thread/async context), so concurrent indexing of other documents isn't
+    # affected and a one-off value never sticks as the process default.
+    with max_concurrency_scope(getattr(opt, "max_concurrency", None)), \
+         llm_params_scope(getattr(opt, "llm_params", None)):
         nodes = parsed.nodes
         strategy = detect_strategy(nodes)
 

pageindex/index/utils.py

@@ -21,47 +21,103 @@ from pprint import pprint
 # `pageindex.config` submodule for those modules.
 from types import SimpleNamespace as _config
 
-from ..config import get_llm_params, get_max_concurrency
+from contextlib import asynccontextmanager
+
+from ..config import get_llm_params, get_max_concurrency, _process_wide_max_concurrency
 from ..tokens import count_tokens  # re-exported for backward compat
 
 logger = logging.getLogger(__name__)
 
 
-# One shared semaphore per event loop, bounding concurrent in-flight LLM calls.
-# Keyed by the loop object (WeakKeyDictionary drops the entry once the loop is
-# closed and garbage-collected) so each asyncio.run() gets its own, correctly
-# loop-bound semaphore. The lock only guards the tiny get-or-create against two
-# threads (each driving its own loop) racing to insert; within a single loop
-# everything is single-threaded, so no lock is needed on the hot path.
-_LLM_SEMAPHORES: "weakref.WeakKeyDictionary" = weakref.WeakKeyDictionary()
-_LLM_SEMAPHORES_LOCK = threading.Lock()
+# TRUE process-wide ceiling on concurrent in-flight LLM calls, shared across
+# EVERY thread and event loop (a plain threading.Semaphore, not an
+# asyncio.Semaphore — those are bound to the loop that created them, so one per
+# loop would let N concurrently-indexing threads each get their own full-size
+# cap and multiply the effective bound by N). Resized lazily when the
+# process-wide default changes; resizing isn't perfectly atomic against
+# in-flight acquires, which is fine since it only happens on an explicit
+# set_max_concurrency() config change, not on the hot path.
+_PROCESS_LLM_SEMAPHORE: threading.Semaphore | None = None
+_PROCESS_LLM_SEMAPHORE_SIZE: int | None = None
+_PROCESS_LLM_SEMAPHORE_LOCK = threading.Lock()
+
+# Per-loop, per-size semaphores for a max_concurrency_scope() override that's
+# narrower than the process ceiling — isolates one call's own subtree to a
+# tighter self-imposed limit without needing to be cross-thread itself (it can
+# never let MORE calls through than the process ceiling above already allows,
+# since both are held simultaneously; see _llm_semaphore). Keyed by (loop, size)
+# rather than just loop so a later scope with a different size in the same loop
+# isn't silently ignored.
+_SCOPED_LLM_SEMAPHORES: "weakref.WeakKeyDictionary" = weakref.WeakKeyDictionary()
+_SCOPED_LLM_SEMAPHORES_LOCK = threading.Lock()
 
 
-def _llm_semaphore() -> asyncio.Semaphore:
-    """Shared per-loop cap on concurrent in-flight LLM calls.
+def _process_ceiling_semaphore() -> threading.Semaphore:
+    global _PROCESS_LLM_SEMAPHORE, _PROCESS_LLM_SEMAPHORE_SIZE
+    size = _process_wide_max_concurrency()
+    with _PROCESS_LLM_SEMAPHORE_LOCK:
+        if _PROCESS_LLM_SEMAPHORE is None or _PROCESS_LLM_SEMAPHORE_SIZE != size:
+            _PROCESS_LLM_SEMAPHORE = threading.Semaphore(size)
+            _PROCESS_LLM_SEMAPHORE_SIZE = size
+        return _PROCESS_LLM_SEMAPHORE
+
+
+def _scoped_llm_semaphore(size: int) -> asyncio.Semaphore:
+    loop = asyncio.get_running_loop()
+    per_loop = _SCOPED_LLM_SEMAPHORES.get(loop)
+    if per_loop is None:
+        with _SCOPED_LLM_SEMAPHORES_LOCK:
+            per_loop = _SCOPED_LLM_SEMAPHORES.get(loop)
+            if per_loop is None:
+                per_loop = {}
+                _SCOPED_LLM_SEMAPHORES[loop] = per_loop
+    sem = per_loop.get(size)
+    if sem is None:
+        with _SCOPED_LLM_SEMAPHORES_LOCK:
+            sem = per_loop.get(size)
+            if sem is None:
+                sem = asyncio.Semaphore(size)
+                per_loop[size] = sem
+    return sem
+
+
+@asynccontextmanager
+async def _llm_semaphore():
+    """Bound concurrent in-flight LLM calls to a TRUE process-wide ceiling,
+    optionally narrowed further by an active max_concurrency_scope() override.
 
     Acquired only around the leaf ``litellm.acompletion`` call in
     ``llm_acompletion`` — the single point every LLM request funnels through —
-    so the cap is a TRUE global bound no matter how deeply the indexing gathers
-    nest (``tree_parser`` → ``process_large_node_recursively`` → …). Bounding at
-    the leaf rather than at each gather call site is also deadlock-free: a parent
-    coroutine awaiting its children holds no slot, so children can always
-    acquire one.
+    so the cap holds no matter how deeply the indexing gathers nest
+    (``tree_parser`` → ``process_large_node_recursively`` → …) AND no matter how
+    many threads are each running their own indexing job concurrently. Bounding
+    at the leaf rather than at each gather call site is also deadlock-free: a
+    parent coroutine awaiting its children holds no slot, so children can
+    always acquire one.
 
-    Sized from ``get_max_concurrency()`` the first time it's needed in a loop, so
-    a per-index ``max_concurrency_scope`` override in effect at that moment is
-    honored. Without this bound a many-node document opens one socket per node at
-    once and exhausts the process file-descriptor limit (Errno 24).
+    The process ceiling (threading.Semaphore, shared cross-thread) is sized from
+    the process-wide default only; a narrower max_concurrency_scope() override
+    is enforced as a second, nested, per-loop restriction — it can only
+    *tighten* the effective cap for its own call tree, never widen it past the
+    ceiling. Without the outer bound a many-node document opens one socket per
+    node at once and exhausts the process file-descriptor limit (Errno 24).
     """
-    loop = asyncio.get_running_loop()
-    sem = _LLM_SEMAPHORES.get(loop)
-    if sem is None:
-        with _LLM_SEMAPHORES_LOCK:
-            sem = _LLM_SEMAPHORES.get(loop)
-            if sem is None:
-                sem = asyncio.Semaphore(get_max_concurrency())
-                _LLM_SEMAPHORES[loop] = sem
-    return sem
+    ceiling_sem = _process_ceiling_semaphore()
+    # threading.Semaphore.acquire() blocks the calling thread, so run it off
+    # the event loop thread — otherwise it would freeze every other coroutine
+    # on this loop while waiting for a slot. release() is non-blocking and
+    # safe to call directly from any thread.
+    await asyncio.to_thread(ceiling_sem.acquire)
+    try:
+        effective = get_max_concurrency()
+        ceiling = _process_wide_max_concurrency()
+        if effective < ceiling:
+            async with _scoped_llm_semaphore(effective):
+                yield
+        else:
+            yield
+    finally:
+        ceiling_sem.release()
 
 
 def llm_completion(model, prompt, chat_history=None, return_finish_reason=False):
@@ -258,7 +314,14 @@ async def generate_node_summary(node, model=None):
 async def generate_summaries_for_structure(structure, model=None):
     nodes = structure_to_list(structure)
     tasks = [generate_node_summary(node, model=model) for node in nodes]
-    summaries = await asyncio.gather(*tasks)
+    # return_exceptions=True: one node's summary failing (e.g. a transient LLM
+    # error) must not abort summarization for the whole document — fall back
+    # to the node's own raw text so retrieval still has something usable.
+    raw_summaries = await asyncio.gather(*tasks, return_exceptions=True)
+    summaries = [
+        node.get('text', '') if isinstance(s, Exception) else s
+        for node, s in zip(nodes, raw_summaries)
+    ]
 
     for node, summary in zip(nodes, summaries):
         node['summary'] = summary
@@ -829,7 +892,13 @@ class ConfigLoader:
 
         self._validate_keys(user_dict)
         merged = {**self._default_dict, **user_dict}
-        return _config(**merged)
+        # Route through IndexConfig so legacy 'yes'/'no' string overrides get
+        # pydantic's bool coercion (a bare 'no' is otherwise a truthy string —
+        # page_index_main's `if opt.if_add_node_summary:` checks would silently
+        # invert the caller's intent).
+        from ..config import IndexConfig
+        validated = IndexConfig(**merged)
+        return _config(**validated.model_dump())
 
 
 def create_node_mapping(tree, include_page_ranges=False, max_page=None):