iatoolkit 2.75.2__tar.gz → 2.75.4__tar.gz

{iatoolkit-2.75.2/src/iatoolkit.egg-info → iatoolkit-2.75.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.75.2
+Version: 2.75.4
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT

{iatoolkit-2.75.2 → iatoolkit-2.75.4}/src/iatoolkit/__init__.py

@@ -3,7 +3,7 @@
 #
 # IAToolkit is open source software.
 
-__version__ = "2.75.2"
+__version__ = "2.75.4"
 
 # Expose main classes and functions at the top level of the package
 

{iatoolkit-2.75.2 → iatoolkit-2.75.4}/src/iatoolkit/services/history_compaction_service.py

@@ -93,7 +93,9 @@ class HistoryCompactionService:
     def _compact(self, company_short_name: str, user_identifier: str, model: str,
                  max_chunks: Optional[int]) -> dict:
         history = self.session_context.get_context_history(company_short_name, user_identifier, model=model) or []
-        budget = self.history_manager.context_budget_tokens(company_short_name, model)
+        # What the recent turns may keep verbatim is a share of the size the
+        # history is compacted toward, not of the whole budget.
+        budget = self.history_manager.compaction_budget_tokens(company_short_name, model)
         plan = self.plan(history, budget)
         if plan is None:
             return {"status": "skipped", "reason": "nothing_to_compact"}

{iatoolkit-2.75.2 → iatoolkit-2.75.4}/src/iatoolkit/services/history_manager_service.py

@@ -46,11 +46,29 @@ class HistoryManagerService:
     #: set its own with `llm.max_context_tokens`.
     DEFAULT_MAX_CONTEXT_TOKENS = 200000
 
-    #: Share of the budget at which a finished turn asks for a background
-    #: compaction, and at which the next turn compacts in line because the
-    #: background one did not land in time (BL-009, Fase E).
+    #: Share of the compaction budget at which a finished turn asks for a
+    #: background compaction, and share of the budget at which the next turn
+    #: compacts in line because the background one did not land in time
+    #: (BL-009, Fase E).
     COMPACTION_TRIGGER_SHARE = 0.70
     INLINE_COMPACTION_SHARE = 0.95
+    #: The size the history is kept near, whatever the budget: the budget stops
+    #: a request from overflowing, this keeps every turn from resending 140k
+    #: tokens (70% of the 200k default) before the first compaction. 64k
+    #: compacted every few SQL turns; with old tool outputs cut, 100k (asks at
+    #: 70k) leaves a long conversation alone. A company may set its own with
+    #: `llm.compaction_budget_tokens`.
+    DEFAULT_COMPACTION_BUDGET_TOKENS = 100000
+
+    #: Tool outputs of the turns older than this many are cut down in the
+    #: stored history: an SQL result or a schema is what the model needs on the
+    #: turn that asked for it, and resending it whole on every later turn was
+    #: most of the history's growth (4k -> 11k tokens on one SQL turn in
+    #: production). The model can call the tool again.
+    KEEP_TOOL_OUTPUT_TURNS = 2
+    #: What is kept of a cut output: its beginning, enough to remember its shape.
+    TRIMMED_TOOL_OUTPUT_CHARS = 1500
+    TRIMMED_TOOL_OUTPUT_MARKER = "[tool output trimmed"
     #: Parts an in-line compaction may summarize: one, the person is waiting.
     INLINE_COMPACTION_MAX_CHUNKS = 1
     #: How long a turn waits for a compaction that is writing the history.
@@ -201,6 +219,7 @@ class HistoryManagerService:
         elif response.get('answer'):
             context_history.append({"role": "assistant", "content": response.get('answer', '')})
 
+        self._trim_old_tool_outputs(context_history)
         self._cap_stored_history(company_short_name, model, context_history)
 
         self.session_context.save_context_history(
@@ -223,7 +242,15 @@ class HistoryManagerService:
         try:
             self.session_context.save_context_tokens(company_short_name, user_identifier, context_tokens, model=model)
             budget = self.context_budget_tokens(company_short_name, model)
-            if context_tokens >= budget * self.COMPACTION_TRIGGER_SHARE:
+            threshold = int(self.compaction_budget_tokens(company_short_name, model) * self.COMPACTION_TRIGGER_SHARE)
+            compact = context_tokens >= threshold
+            # One line per turn: how a conversation grows, and when it asked for a compaction.
+            logging.info(
+                f"Context of {company_short_name}/{user_identifier} (model={model}): {context_tokens} tokens, "
+                f"{context_tokens * 100 // max(budget, 1)}% of the {budget} budget; compaction at {threshold}"
+                f"{' -> requested' if compact else ''}."
+            )
+            if compact:
                 self.compaction_trigger.trigger(company_short_name, user_identifier, model)
         except Exception as e:
             # A compaction that is not asked for now is asked for on the next turn.
@@ -314,8 +341,18 @@ class HistoryManagerService:
         cap = self._company_context_cap(company_short_name) or self.DEFAULT_MAX_CONTEXT_TOKENS
         return min(budget, cap)
 
+    def compaction_budget_tokens(self, company_short_name: str, model: Optional[str]) -> int:
+        """The size the history is compacted toward: never above the budget."""
+        target = (self._company_llm_setting(company_short_name, "compaction_budget_tokens")
+                  or self.DEFAULT_COMPACTION_BUDGET_TOKENS)
+        return min(target, self.context_budget_tokens(company_short_name, model))
+
     def _company_context_cap(self, company_short_name: str) -> int:
         """`llm.max_context_tokens` in company.yaml; 0 when not set (the platform default applies)."""
+        return self._company_llm_setting(company_short_name, "max_context_tokens")
+
+    def _company_llm_setting(self, company_short_name: str, key: str) -> int:
+        """A positive integer under `llm` in company.yaml; 0 when not set."""
         if self.configuration_service is None:
             return 0
         try:
@@ -324,7 +361,7 @@ class HistoryManagerService:
             return 0
         if not isinstance(llm_config, dict):
             return 0
-        return self._positive_int(llm_config.get("max_context_tokens"))
+        return self._positive_int(llm_config.get(key))
 
     @staticmethod
     def _positive_int(value) -> int:
@@ -349,6 +386,40 @@ class HistoryManagerService:
             and not cls.is_summary_message(message)
         ]
 
+    @classmethod
+    def _trim_old_tool_outputs(cls, context_history: list):
+        """Cuts the long tool outputs of all but the last KEEP_TOOL_OUTPUT_TURNS turns, in place.
+
+        Each output keeps its message (a tool call without its output is a
+        history every provider rejects) and its beginning, plus a note naming the
+        tool so the model knows it can call it again. An output is cut once: the
+        stored history then stays the same from turn to turn, which is what a
+        provider's prompt cache reuses.
+        """
+        turn_starts = cls.turn_starts(context_history)
+        if len(turn_starts) <= cls.KEEP_TOOL_OUTPUT_TURNS:
+            return
+        boundary = turn_starts[-cls.KEEP_TOOL_OUTPUT_TURNS]
+        tool_names = {}
+        for message in context_history[:boundary]:
+            for call in (message.get("tool_calls") or []) if isinstance(message, dict) else []:
+                if isinstance(call, dict):
+                    tool_names[call.get("id")] = (call.get("function") or {}).get("name")
+        for message in context_history[:boundary]:
+            if not isinstance(message, dict) or message.get("type") != "function_call_output":
+                continue
+            output = message.get("output")
+            if not isinstance(output, str) or len(output) <= cls.TRIMMED_TOOL_OUTPUT_CHARS:
+                continue
+            if cls.TRIMMED_TOOL_OUTPUT_MARKER in output[cls.TRIMMED_TOOL_OUTPUT_CHARS:]:
+                continue
+            tool = tool_names.get(message.get("call_id")) or "the tool"
+            message["output"] = (
+                f"{output[:cls.TRIMMED_TOOL_OUTPUT_CHARS]}\n"
+                f"{cls.TRIMMED_TOOL_OUTPUT_MARKER}: {len(output)} characters, from an earlier turn. "
+                f"Call {tool} again if you need all of it.]"
+            )
+
     def _cap_stored_history(self, company_short_name: str, model: Optional[str], context_history: list):
         """Keeps the stored history under STORED_HISTORY_MAX_BUDGETS budgets, whole turns at a time."""
         cap = self.context_budget_tokens(company_short_name, model) * self.STORED_HISTORY_MAX_BUDGETS

{iatoolkit-2.75.2 → iatoolkit-2.75.4}/src/iatoolkit/views/base_login_view.py

@@ -55,11 +55,13 @@ class BaseLoginView(MethodView):
                            target_url: str,
                            redeem_token: str = None):
         """
-        Centralized logic to decide between the fast path and the slow path.
+        Starts the session's context and renders the chat. `target_url` (the old
+        loading shell's iframe) is no longer used: since BL-009 starting the context
+        calls no model, so there is nothing to wait for behind onboarding cards.
         """
         self._trigger_warmup(company_short_name, "post_login")
 
-        # --- Get the company branding and onboarding_cards
+        # --- Get the company branding and onboarding_cards (the chat's help button shows them)
         branding_data = self.branding_service.get_company_branding(company_short_name)
         onboarding_cards = self.config_service.get_configuration(company_short_name, 'onboarding_cards')
 
@@ -67,46 +69,39 @@ class BaseLoginView(MethodView):
         prep_result = self.query_service.prepare_context(
             company_short_name=company_short_name, user_identifier=user_identifier
         )
-
         if prep_result.get('rebuild_needed'):
-            # --- SLOW PATH: Render the loading shell ---
-            return render_template(
-                "onboarding_shell.html",
-                company_short_name=company_short_name,
-                user_identifier=user_identifier,
-                iframe_src_url=target_url,
-                branding=branding_data,
-                onboarding_cards=onboarding_cards
+            # Starts the history with the prepared context, in the session only.
+            self.query_service.set_context_for_llm(
+                company_short_name=company_short_name, user_identifier=user_identifier
             )
-        else:
-            # --- FAST PATH: Render the chat page directly ---
-            # LLM configuration: default model and availables
-            default_llm_model, available_llm_models = self.config_service.get_llm_configuration(company_short_name)
-            default_llm_model, available_llm_models = access.filter_llm_models(
-                company_short_name, user_identifier, default_llm_model, available_llm_models)
-            llm_request_defaults = self.config_service.get_llm_request_defaults(company_short_name) or {}
-            llm_default_reasoning_effort = str(
-                ((llm_request_defaults.get("reasoning") or {}).get("effort")) or ""
-            ).strip().lower()
 
-            prompts = self.prompt_service.get_prompts(company_short_name, user_identifier=user_identifier)
+        # LLM configuration: default model and availables
+        default_llm_model, available_llm_models = self.config_service.get_llm_configuration(company_short_name)
+        default_llm_model, available_llm_models = access.filter_llm_models(
+            company_short_name, user_identifier, default_llm_model, available_llm_models)
+        llm_request_defaults = self.config_service.get_llm_request_defaults(company_short_name) or {}
+        llm_default_reasoning_effort = str(
+            ((llm_request_defaults.get("reasoning") or {}).get("effort")) or ""
+        ).strip().lower()
+
+        prompts = self.prompt_service.get_prompts(company_short_name, user_identifier=user_identifier)
 
-            # Get the entire 'js_messages' block in the correct language.
-            js_translations = self.i18n_service.get_translation_block('js_messages')
+        # Get the entire 'js_messages' block in the correct language.
+        js_translations = self.i18n_service.get_translation_block('js_messages')
 
-            return render_template(
-                "chat.html",
-                company_short_name=company_short_name,
-                user_identifier=user_identifier,
-                prompts=prompts,
-                branding=branding_data,
-                onboarding_cards=onboarding_cards,
-                js_translations=js_translations,
-                redeem_token=redeem_token,
-                llm_default_model=default_llm_model,
-                llm_available_models = available_llm_models,
-                llm_default_reasoning_effort=llm_default_reasoning_effort,
-                )
+        return render_template(
+            "chat.html",
+            company_short_name=company_short_name,
+            user_identifier=user_identifier,
+            prompts=prompts,
+            branding=branding_data,
+            onboarding_cards=onboarding_cards,
+            js_translations=js_translations,
+            redeem_token=redeem_token,
+            llm_default_model=default_llm_model,
+            llm_available_models = available_llm_models,
+            llm_default_reasoning_effort=llm_default_reasoning_effort,
+        )
 
     def _trigger_warmup(self, company_short_name: str, trigger: str):
         try:

{iatoolkit-2.75.2 → iatoolkit-2.75.4/src/iatoolkit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.75.2
+Version: 2.75.4
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT

{iatoolkit-2.75.2 → iatoolkit-2.75.4}/src/iatoolkit.egg-info/SOURCES.txt

@@ -236,7 +236,6 @@ src/iatoolkit/templates/error.html
 src/iatoolkit/templates/forgot_password.html
 src/iatoolkit/templates/home_hosted_default.html
 src/iatoolkit/templates/login_global.html
-src/iatoolkit/templates/onboarding_shell.html
 src/iatoolkit/templates/signup.html
 src/iatoolkit/templates/wiki_page.html
 src/iatoolkit/templates/memory/wiki_schema.md

iatoolkit-2.75.2/src/iatoolkit/templates/onboarding_shell.html

@@ -1,106 +0,0 @@
-{% extends "base.html" %}
-
-{% block title %}Iniciando {{ branding.name }} IA...{% endblock %}
-
-{% block styles %}
-{% endblock %}
-
-{% block content %}
-
-<link rel="stylesheet" href="{{ url_for('static', filename='styles/llm_output.css', _external=True) }}?v=6">
-<link rel="stylesheet" href="{{ url_for('static', filename='styles/onboarding.css', _external=True) }}?v=6">
-
-<style>
-  {% if branding and branding.css_variables %}
-    {{ branding.css_variables|safe }}
-  {% endif %}
-
-  {# 2. Ahora definimos las reglas que usan esas variables #}
-  .onboarding-shell-root { position: relative; height: 100vh; }
-  .onboarding-shell-root #loader-wrapper { position: absolute; inset: 0; background: #f4f7f6; z-index: 1000; display:flex; align-items:center; justify-content:center; padding:20px; box-sizing:border-box; transition: opacity .5s; }
-  .onboarding-shell-root .ob-stack { width:100%; max-width:520px; display:flex; flex-direction:column; gap:16px; }
-  .onboarding-shell-root .ob-loading-band { display:flex; align-items:center; justify-content:center; gap:12px; padding:12px 14px; background:#fff; border-radius:12px; box-shadow:0 4px 20px rgba(0,0,0,.08); }
-  .onboarding-shell-root .spinner { width:28px; height:28px; border:4px solid rgba(0,0,0,.1); border-top-color: var(--brand-primary-color, #FF5100); border-radius:50%; animation: spin 1s linear infinite; }
-  @keyframes spin { to { transform: rotate(360deg); } }
-  .onboarding-shell-root #content-container { width:100%; height:100%; }
-  .onboarding-shell-root #content-container iframe { width:100%; height:100%; border:none; }
-
-  /* Forzar colores de marca (sube especificidad y evita overrides) */
-  .onboarding-shell-root .ob-brand-header { color: var(--brand-secondary-color, #06326B) !important; }
-  .onboarding-shell-root .ob-brand-header .brand-name { color: var(--brand-primary-color, #FF5100) !important; }
-  .onboarding-shell-root .ob-icon { color: var(--brand-primary-color, #FF5100) !important; }
-  .onboarding-shell-root .ob-btn { background-color: var(--brand-secondary-color, #06326B) !important; color: var(--brand-text-on-secondary, #FFFFFF) !important; border:none !important; }
-  .onboarding-shell-root .ob-dots > div.active { background-color: var(--brand-primary-color, #FF5100) !important; }
-</style>
-
-<div class="onboarding-shell-root ob-root">
-  <div id="loader-wrapper">
-    <div class="ob-stack">
-      <h1 id="ob-brand-header" class="ob-brand-header">
-        <span class="brand-name text-brand-primary">{{ branding.name | default('IAToolkit') }}</span>
-      </h1>
-
-      <div id="card-container" class="ob-card">
-        <div id="card-icon" class="ob-icon"><i class="fas fa-lightbulb"></i></div>
-        <h3 id="card-title" class="ob-title">Título de la Tarjeta</h3>
-        <p id="card-text" class="ob-text">Descripción de la tarjeta de capacitación.</p>
-        <p id="card-example" class="ob-example"></p>
-
-        <div class="ob-nav">
-          <button id="prev-card" class="ob-btn btn-branded-primary" aria-label="Anterior">
-            <i class="fas fa-chevron-left"></i>
-          </button>
-          <div id="progress-dots" class="ob-dots"></div>
-          <button id="next-card" class="ob-btn btn-branded-primary" aria-label="Siguiente">
-            <i class="fas fa-chevron-right"></i>
-          </button>
-        </div>
-      </div>
-
-      <div id="loading-status" class="ob-loading-band">
-        <div class="spinner"></div>
-        <p>{{ t('ui.chat.init_context') }} </p>
-      </div>
-    </div>
-  </div>
-
-  <div id="content-container"></div>
-</div>
-{% endblock %}
-
-{% block scripts %}
-<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.4/css/all.min.css">
-
-<script src="{{ url_for('static', filename='js/chat_onboarding_button.js', _external=True) }}"></script>
-<script>
-  (function() {
-    const cardsData = {{ onboarding_cards | tojson }};
-    const iframeSrc = "{{ iframe_src_url }}";
-
-    if (!window.initOnboarding) {
-      console.error('initOnboarding no está disponible. Verifica la carga de chat_onboarding.js');
-      return;
-    }
-
-    const onboarding = initOnboarding({
-      mode: 'shell',
-      cards: cardsData,
-      ui: {
-        icon: '#card-icon',
-        title: '#card-title',
-        text: '#card-text',
-          example: '#card-example',
-        dots: '#progress-dots',
-        prev: '#prev-card',
-        next: '#next-card',
-        loader: '#loader-wrapper',
-        container: '#content-container'
-      },
-      autoRotateMs: 5000,
-      shell: { iframeSrc: iframeSrc }
-    });
-
-    onboarding.start();
-  })();
-</script>
-{% endblock %}