cat-stack 1.5.0__tar.gz → 1.6.0__tar.gz

{cat_stack-1.5.0 → cat_stack-1.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cat-stack
-Version: 1.5.0
+Version: 1.6.0
 Summary: Domain-agnostic text, image, PDF, and DOCX classification engine powered by LLMs
 Project-URL: Documentation, https://github.com/chrissoria/cat-stack#readme
 Project-URL: Issues, https://github.com/chrissoria/cat-stack/issues
@@ -181,6 +181,57 @@ All providers use the same `(model_name, provider, api_key)` tuple format. Provi
 - **Embedding similarity** tiebreaker for ensemble consensus ties
 - **Pilot test** — validate classifications on a small sample before committing to the full run
 
+## Future work / contributions welcome
+
+The following items are tracked but not yet implemented. PRs welcome —
+each entry includes the scope I'd suggest if someone wants to pick it up.
+
+- **Standalone SambaNova provider.** Currently SambaNova-hosted models
+  are reachable through the HuggingFace router suffix
+  (`meta-llama/...:sambanova`), but there's no direct
+  `provider="sambanova"` path that talks to SambaNova's own
+  OpenAI-compatible endpoint. Wiring it up means a new
+  `PROVIDER_CONFIG` entry, the right base URL
+  (`https://api.sambanova.ai/v1`), token-detection rules in
+  `detect_provider`, and a smoke test against one of their cheap
+  models (e.g. `Meta-Llama-3.1-8B-Instruct`).
+
+- **Consolidate HuggingFace-suffix dispatch.** The strings
+  `"huggingface"` and `"huggingface-together"` are currently
+  hardcoded in ~30 dispatch sites across
+  `pdf_functions.py` / `image_functions.py` /
+  `text_functions_ensemble.py` / `_chunked.py`. Adding a new router
+  suffix (e.g. `huggingface-fireworks`) means updating every one of
+  them. The cleaner refactor is a single
+  `_is_openai_compatible(model_source)` helper that matches anything
+  starting with `huggingface` plus the static list
+  (openai/perplexity/xai). Same shape as our existing
+  `_sanitize_google_schema` helper. Touches a lot of sites but each
+  edit is mechanical.
+
+- **Meta-LLM "Senate VP" tiebreaker + batch_mode support for
+  `embedding_tiebreaker`.** The existing `embedding_tiebreaker=True`
+  resolves true 50/50 ties via centroid similarity, but only in
+  synchronous ensemble mode. Two related extensions: (a) a meta-LLM
+  tie-breaker that invokes a separate model on tied rows
+  (`tie_break="meta_model"` with a configurable model); (b) extend
+  the existing centroid tiebreaker to work inside `batch_mode=True`
+  by running it after the batch results come back, before
+  `build_output_dataframes`. The infrastructure for both is already
+  in `_tiebreaker.py`; the meta-LLM variant would be a new resolver
+  function called from `resolve_ties_with_centroids`.
+
+- **Schema-permafail retry short-circuit.** When a model's
+  classification permanently fails schema validation across all
+  available retry budgets, the framework keeps spending API calls.
+  A short-circuit that detects "this model + this input is producing
+  the same invalid output N times in a row" and bails out early
+  would save quota. Scope was narrowed earlier (after the
+  HF-SMALL-MODEL fix reduced the wasted-retries surface area), so
+  there's a real risk this stays low-value; recommend writing the
+  detection metric first, instrumenting an actual run, and only
+  building the short-circuit if the metric says it would have helped.
+
 ## License
 
 GPL-3.0-or-later

{cat_stack-1.5.0 → cat_stack-1.6.0}/README.md

@@ -145,6 +145,57 @@ All providers use the same `(model_name, provider, api_key)` tuple format. Provi
 - **Embedding similarity** tiebreaker for ensemble consensus ties
 - **Pilot test** — validate classifications on a small sample before committing to the full run
 
+## Future work / contributions welcome
+
+The following items are tracked but not yet implemented. PRs welcome —
+each entry includes the scope I'd suggest if someone wants to pick it up.
+
+- **Standalone SambaNova provider.** Currently SambaNova-hosted models
+  are reachable through the HuggingFace router suffix
+  (`meta-llama/...:sambanova`), but there's no direct
+  `provider="sambanova"` path that talks to SambaNova's own
+  OpenAI-compatible endpoint. Wiring it up means a new
+  `PROVIDER_CONFIG` entry, the right base URL
+  (`https://api.sambanova.ai/v1`), token-detection rules in
+  `detect_provider`, and a smoke test against one of their cheap
+  models (e.g. `Meta-Llama-3.1-8B-Instruct`).
+
+- **Consolidate HuggingFace-suffix dispatch.** The strings
+  `"huggingface"` and `"huggingface-together"` are currently
+  hardcoded in ~30 dispatch sites across
+  `pdf_functions.py` / `image_functions.py` /
+  `text_functions_ensemble.py` / `_chunked.py`. Adding a new router
+  suffix (e.g. `huggingface-fireworks`) means updating every one of
+  them. The cleaner refactor is a single
+  `_is_openai_compatible(model_source)` helper that matches anything
+  starting with `huggingface` plus the static list
+  (openai/perplexity/xai). Same shape as our existing
+  `_sanitize_google_schema` helper. Touches a lot of sites but each
+  edit is mechanical.
+
+- **Meta-LLM "Senate VP" tiebreaker + batch_mode support for
+  `embedding_tiebreaker`.** The existing `embedding_tiebreaker=True`
+  resolves true 50/50 ties via centroid similarity, but only in
+  synchronous ensemble mode. Two related extensions: (a) a meta-LLM
+  tie-breaker that invokes a separate model on tied rows
+  (`tie_break="meta_model"` with a configurable model); (b) extend
+  the existing centroid tiebreaker to work inside `batch_mode=True`
+  by running it after the batch results come back, before
+  `build_output_dataframes`. The infrastructure for both is already
+  in `_tiebreaker.py`; the meta-LLM variant would be a new resolver
+  function called from `resolve_ties_with_centroids`.
+
+- **Schema-permafail retry short-circuit.** When a model's
+  classification permanently fails schema validation across all
+  available retry budgets, the framework keeps spending API calls.
+  A short-circuit that detects "this model + this input is producing
+  the same invalid output N times in a row" and bails out early
+  would save quota. Scope was narrowed earlier (after the
+  HF-SMALL-MODEL fix reduced the wasted-retries surface area), so
+  there's a real risk this stays low-value; recommend writing the
+  detection metric first, instrumenting an actual run, and only
+  building the short-circuit if the metric says it would have helped.
+
 ## License
 
 GPL-3.0-or-later

{cat_stack-1.5.0 → cat_stack-1.6.0}/src/catstack/__about__.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-present Christopher Soria <chrissoria@berkeley.edu>
 #
 # SPDX-License-Identifier: GPL-3.0-or-later
-__version__ = "1.5.0"
+__version__ = "1.6.0"
 __author__ = "Chris Soria"
 __email__ = "chrissoria@berkeley.edu"
 __title__ = "cat-stack"

{cat_stack-1.5.0 → cat_stack-1.6.0}/src/catstack/_batch.py

@@ -102,6 +102,58 @@ class BatchJobFailedError(RuntimeError):
     pass
 
 
+def _inspect_anthropic_terminal_state(status_data: dict, job_id: str) -> None:
+    """Inspect an Anthropic batch in `processing_status == "ended"`.
+
+    Anthropic uses a single terminal state ("ended") for all outcomes —
+    fully succeeded, fully errored, fully canceled, fully expired, or
+    any mix. The polling code treats "ended" as success and returns
+    status_data; per-request errors get surfaced at parse time. That
+    works for the mixed case but is misleading when 0/N requests
+    succeeded: the caller silently gets a DataFrame of all-error rows
+    with no clear signal that the whole batch was dead.
+
+    This helper raises the appropriate exception when the batch is
+    *uniformly* failed/canceled/expired, and prints a warning for the
+    partial-failure case. Returns silently when the batch has any
+    successes (combined with per-row errors from parse layer).
+    """
+    counts = status_data.get("request_counts", {})
+    succeeded = counts.get("succeeded", 0)
+    errored = counts.get("errored", 0)
+    canceled = counts.get("canceled", 0)
+    expired = counts.get("expired", 0)
+    total = succeeded + errored + canceled + expired
+
+    if total == 0:
+        return
+
+    if succeeded == 0:
+        if canceled == total:
+            raise BatchJobExpiredError(
+                f"Anthropic batch '{job_id}' was canceled (canceled={canceled}). "
+                f"Job ID saved above — check provider dashboard for details."
+            )
+        if expired == total:
+            raise BatchJobExpiredError(
+                f"Anthropic batch '{job_id}' expired before any requests succeeded "
+                f"(expired={expired})."
+            )
+        raise BatchJobFailedError(
+            f"Anthropic batch '{job_id}' ended with 0/{total} requests succeeded "
+            f"(errored={errored}, canceled={canceled}, expired={expired}). "
+            f"Check the provider dashboard for the error details."
+        )
+
+    if errored or canceled or expired:
+        print(
+            f"  [batch] Anthropic batch '{job_id}' partial: "
+            f"succeeded={succeeded}, errored={errored}, "
+            f"canceled={canceled}, expired={expired}. "
+            f"Errored rows will appear as failures in the output DataFrame."
+        )
+
+
 # =============================================================================
 # Auth headers
 # =============================================================================
@@ -451,6 +503,8 @@ def _poll_batch_job(
                     f"Batch job '{job_id}' failed (state: {state}). "
                     f"Check the provider dashboard for details."
                 )
+            if provider == "anthropic":
+                _inspect_anthropic_terminal_state(status_data, job_id)
             return status_data
 
         time.sleep(interval)
@@ -718,6 +772,7 @@ def _run_one_batch_job(
             stepback_insights=stepback_insights,
             model_name=model,
             multi_label=prompt_params.get("multi_label", True),
+            system_prompt=prompt_params.get("system_prompt", ""),
         )
 
         payload = client._build_payload(
@@ -816,6 +871,7 @@ def _run_one_sync_model(
             stepback_insights=prompt_params.get("stepback_insights", {}),
             model_name=model,
             multi_label=prompt_params.get("multi_label", True),
+            system_prompt=prompt_params.get("system_prompt", ""),
         )
         try:
             raw, err = client.complete(
@@ -998,7 +1054,20 @@ def run_batch_ensemble_classify(
     with ThreadPoolExecutor(max_workers=len(model_configs)) as executor:
         futures = {executor.submit(_run_cfg, cfg): cfg for cfg in model_configs}
         for future in as_completed(futures):
-            model_key, result = future.result()
+            cfg = futures[future]
+            model_key = cfg["sanitized_name"]
+            try:
+                _, result = future.result()
+            except Exception as e:
+                print(
+                    f"\n[batch ensemble] Model '{cfg['model']}' ({cfg['provider']}) "
+                    f"failed: {type(e).__name__}: {e}"
+                )
+                print(
+                    f"  Other models will still complete; "
+                    f"this model's column will be empty."
+                )
+                result = {}
             all_model_results[model_key] = result
 
     all_results = []
@@ -1313,7 +1382,20 @@ def run_batch_ensemble_summarize(
     with ThreadPoolExecutor(max_workers=len(model_configs)) as executor:
         futures = {executor.submit(_run_cfg, cfg): cfg for cfg in model_configs}
         for future in as_completed(futures):
-            model_key, result = future.result()
+            cfg = futures[future]
+            model_key = cfg["sanitized_name"]
+            try:
+                _, result = future.result()
+            except Exception as e:
+                print(
+                    f"\n[batch ensemble] Model '{cfg['model']}' ({cfg['provider']}) "
+                    f"failed: {type(e).__name__}: {e}"
+                )
+                print(
+                    f"  Other models will still complete; "
+                    f"this model's column will be empty."
+                )
+                result = {}
             all_model_results[model_key] = result
 
     model_names = [cfg["sanitized_name"] for cfg in model_configs]

{cat_stack-1.5.0 → cat_stack-1.6.0}/src/catstack/_formatter.py

@@ -42,31 +42,26 @@ def _check_dependencies():
         )
 
 
-def _ensure_dependencies(verbose: bool = True) -> bool:
-    """Ensure formatter Python dependencies are installed.
-
-    Tries to import torch/transformers/accelerate. If any are missing,
-    auto-installs them via pip after printing a clear warning about the
-    download size (~1.5 GB total). Returns True on success, False on
-    install failure.
-    """
+def _check_dependencies_installed() -> bool:
+    """Pure check — returns True if all formatter deps import successfully.
+    No side effects, no install attempts."""
     try:
         import torch  # noqa: F401
         import transformers  # noqa: F401
         import accelerate  # noqa: F401
         return True
     except ImportError:
-        pass
+        return False
 
-    if verbose:
-        print(
-            "\n[CatLLM] JSON formatter dependencies (transformers, torch, "
-            "accelerate)\n"
-            "  are not installed in this Python environment. Installing now\n"
-            "  (~1.5 GB download; one-time). To skip this and disable the\n"
-            "  formatter, pass json_formatter=False."
-        )
 
+def _install_dependencies(verbose: bool = True) -> bool:
+    """Run `pip install` for the formatter deps. Caller is responsible for
+    obtaining user consent before calling this — it does not prompt.
+
+    Returns True if deps are importable after install, False otherwise.
+    """
+    if verbose:
+        print("[CatLLM] Installing formatter dependencies (~1.5 GB)…")
     import subprocess
     try:
         subprocess.check_call(
@@ -77,19 +72,100 @@ def _ensure_dependencies(verbose: bool = True) -> bool:
         if verbose:
             print(
                 f"[CatLLM] Failed to install formatter dependencies ({e}).\n"
-                "  Install manually: pip install 'cat-llm[formatter]'"
+                "  Install manually: pip install 'cat-stack[formatter]'"
             )
         return False
+    return _check_dependencies_installed()
+
+
+def _prompt_formatter_consent(model_label: str = "the current model") -> str:
+    """Interactive consent prompt for the JSON formatter fallback.
+
+    Two paths depending on whether the formatter dependencies are already
+    installed:
+      - Deps installed: asks whether to load the ~1 GB formatter model.
+      - Deps missing:   asks whether to download deps (~1.5 GB) AND load.
+
+    Non-TTY contexts (CI, batch scripts, headless notebooks): prints a
+    one-time suggestion and returns "declined" without blocking on input.
+
+    Returns "approved" or "declined". On approval with deps missing,
+    also installs the deps before returning.
+    """
+    deps_installed = _check_dependencies_installed()
+
+    if not sys.stdin.isatty():
+        if deps_installed:
+            print(
+                f"\n[CatLLM] Malformed JSON from {model_label}. The JSON "
+                "formatter could recover this — pass json_formatter=True "
+                "to enable, or json_formatter=False to silence this suggestion."
+            )
+        else:
+            print(
+                f"\n[CatLLM] Malformed JSON from {model_label}. The JSON "
+                "formatter could recover, but its deps (~1.5 GB) aren't "
+                "installed. Run `pip install cat-stack[formatter]` and pass "
+                "json_formatter=True to enable, or json_formatter=False to "
+                "silence this suggestion."
+            )
+        return "declined"
+
+    if deps_installed:
+        prompt = (
+            f"\n[CatLLM] {model_label} produced malformed JSON on the first row.\n"
+            "  The JSON formatter can re-format the model's prose output\n"
+            "  into valid catstack JSON for this and subsequent rows.\n"
+            "    Cost: ~1 GB RAM (one-time load).\n"
+            "  Use the formatter for this run? (Y/n): "
+        )
+    else:
+        prompt = (
+            f"\n[CatLLM] {model_label} produced malformed JSON on the first row.\n"
+            "  The JSON formatter can re-format the model's prose output\n"
+            "  into valid catstack JSON for this and subsequent rows.\n"
+            "    Cost: ~1.5 GB download (transformers + torch + accelerate)\n"
+            "         + ~1 GB RAM (one-time load).\n"
+            "  Download deps and use the formatter? (Y/n): "
+        )
 
-    # Verify import works now
     try:
-        import torch  # noqa: F401
-        import transformers  # noqa: F401
+        answer = input(prompt).strip().lower()
+    except (EOFError, KeyboardInterrupt):
+        print("\n[CatLLM] No input received — skipping formatter.")
+        return "declined"
+
+    if answer in ("", "y", "yes"):
+        if not deps_installed:
+            if not _install_dependencies(verbose=True):
+                print("[CatLLM] Continuing without formatter.")
+                return "declined"
+        return "approved"
+    print("[CatLLM] Continuing without formatter.")
+    return "declined"
+
+
+def _ensure_dependencies(verbose: bool = True) -> bool:
+    """Back-compat: ensure deps are installed, auto-installing if missing.
+
+    Still used by the explicit `json_formatter=True` path where the user
+    has already implicitly consented by passing True. The new
+    `json_formatter=None` ("auto") path uses `_prompt_formatter_consent`
+    plus `_install_dependencies` directly so the install requires an
+    explicit yes.
+    """
+    if _check_dependencies_installed():
         return True
-    except ImportError as e:
-        if verbose:
-            print(f"[CatLLM] Formatter deps installed but import failed: {e}")
-        return False
+
+    if verbose:
+        print(
+            "\n[CatLLM] JSON formatter dependencies (transformers, torch, "
+            "accelerate)\n"
+            "  are not installed. Installing now (~1.5 GB download; one-time).\n"
+            "  To skip this and disable the formatter, pass json_formatter=False."
+        )
+
+    return _install_dependencies(verbose=verbose)
 
 
 def _is_model_cached() -> bool: