Deploy: bandit/RAG flags(off)+knowledge mode, RAG/ChromaDB wiring, Vanguard decline tone, analyst commentary (Analysis heading + numbered blue points), no-dup-table, bigger HTML export, larger charts + hbar axis-label fix, JSON-leak fixes

vivek/backend/combined_prompt.py

@@ -386,7 +386,7 @@ Data grounding (STRICT — the widget must mirror your <RESPONSE>):
 
 Honoring an explicitly requested chart type:
 - If the user asks for a SPECIFIC chart type (e.g. "as a candlestick", "show a sankey", "pie chart") and it IS one of the supported kinds above, you MUST use exactly that kind — do not substitute another.
-- If the requested chart type is NOT in the supported kinds (e.g. 3D surface, map/choropleth, gantt, renko, marimekko, etc.), DO NOT silently render a different chart. Instead: in <RESPONSE>, say plainly that you cannot render that specific chart type yet, then name 1-3 supported kinds that would fit their data well and ask if they'd like one of those. In that case return <WIDGET></WIDGET> (empty) — wait for the user to confirm before rendering an alternative.
+- If the requested chart type is NOT in the supported kinds (e.g. 3D surface, map/choropleth, gantt, renko, marimekko, etc.), DO NOT silently render a different chart. Respond in a warm, professional, enterprise tone on behalf of the product "Vanguard" — briefly and politely explain that Vanguard's visualization suite does not currently support that chart type, then helpfully suggest 1-3 supported chart kinds that would present their data well and invite them to choose one. Return <WIDGET></WIDGET> (empty) and wait for them to pick. Do NOT use blunt or technical wording like "not supported in this system", "broken", "incorrect output", or "I cannot render" — keep it courteous and reassuring (e.g. "Vanguard doesn't support 3D surface charts just yet — but I'd be glad to show this as a 3D scatter, treemap, or heatmap instead. Which would you prefer?").
 - Never pretend an unsupported type is supported, and never relabel a different chart as the requested type.
 
 Interactivity:
@@ -396,9 +396,17 @@ Interactivity:
 - If something cannot be expressed with the supported block types (e.g. a playable game, live sliders), DO NOT invent HTML — return <WIDGET></WIDGET> (empty) and explain in <RESPONSE> instead. Never dump raw index arrays like `[0,1,2]`.
 - For photographs, diagrams, or icons, use the `image` block with a valid https:// URL or a data: URI. Combine `image` with `text`, `chart`, and `table` blocks as needed.
 
-Dynamic layout (JSON) — avoid static, single-block dashboards:
-- Shape `layout` like a short story: context first, then metrics, then detail, then actions. Mix block types (text, kpi_row, chart, table, image, action_row) whenever it improves scanning; do not default to one lonely chart if KPIs or a sentence of framing would help.
-- Use `action_row` for obvious follow-up intents; keep blocks ordered top-to-bottom by importance so the widget feels purposeful, not generic.
+Dynamic layout (JSON) — informative, not padded:
+- ALWAYS provide proper DATA-ANALYST commentary with a chart. Put it as a `text` block in the widget BELOW the chart. The block's content MUST start with the heading line "### Analysis" then 3-4 NUMBERED points (each line: "1. ", "2. ", ...). Wrap the KEY figures/labels in **bold** (the renderer highlights bold in blue). Example content:
+"### Analysis
+1. **Data Center** at **$47.5B** is ~**78%** of total FY2024 revenue — the clear dominant segment.
+2. It grew ~**3.2x YoY** (from **$15.0B**), driven by AI/HPC demand.
+3. **Gaming** is a distant second at **$10.4B**; every other segment is under **$2B**.
+4. Takeaway: revenue is heavily **concentrated in AI infrastructure**."
+  Write like a financial/data analyst — specific and quantitative. Across the points cover: leader & laggard with values; the key gap/ratio/share/growth rate; the trend/notable change; any outlier or concentration; and a one-line "so what" takeaway. Use real numbers. NEVER write filler like "this chart shows the data" or "as you can see" — every point must carry a concrete insight. Always 3-4 points (not one). Never show a bare, unexplained chart.
+- Add a `table` when it genuinely HELPS the reader — e.g. exact reference figures, extra columns/notes, or a breakdown the chart can't show. Do NOT add a table that merely restates a simple chart's few values with nothing extra (that is clutter); use one when it adds real detail.
+- You MAY add a few KPI tiles for headline numbers when useful. Mix block types only when each adds value, and keep blocks ordered top-to-bottom by importance.
+- Use `action_row` for obvious follow-up intents so the widget feels purposeful, not generic.
 """
 
 
@@ -411,23 +419,41 @@ _REGISTRY_JSON_PATH = Path(__file__).resolve().parent.parent / "frontend-vue" /
 
 _JSON_RULE_PREAMBLE = """WIDGET JSON SCHEMA MODE (WIDGET_MODE=json):
 - WIDGET WARRANT (decide for yourself — there is no keyword trigger):
-  Include a NON-EMPTY widget ONLY IF both are true: (1) a visual materially helps (a comparison of
-  several numbers, a trend, a breakdown/share, a flow, or a matrix), AND (2) you can populate it with
-  real values from your <RESPONSE> using the supported block types below. If you cannot build it from
-  the allowed blocks (or there is no concrete data), return <WIDGET></WIDGET> (empty) and say so in
-  <RESPONSE>. Do NOT chart a single number, a definition, a yes/no, or an opinion.
+  Include a NON-EMPTY widget when a visual materially helps (a comparison of several numbers, a trend,
+  a breakdown/share, a flow, or a matrix) and you can populate it with concrete values using the
+  supported block types below (see the data rule). Do NOT chart a single number, a definition, a
+  yes/no, or an opinion.
 - The content inside <WIDGET> MUST be valid JSON (no markdown fences, no comments).
 - Root object: { "version": "1.0", "layout": [ ... ] }
 - layout is an ordered array of blocks (top-to-bottom).
 - Supported block types ONLY (do not invent new ones):"""
 
-_JSON_RULE_TRAILER = """
-NO FABRICATED DATA (most important — applies to <RESPONSE> and <WIDGET>):
-- If you do NOT actually have the real figures to answer (e.g. live or historical stock prices, a fund's daily returns, exact financials you are not sure of), DO NOT invent, estimate, or use "illustrative"/"mock"/"example"/"est." data, and DO NOT draw a widget.
-- In that case: say plainly in <RESPONSE> that you do not have that specific data, suggest what the user could provide or ask instead, and return <WIDGET></WIDGET> (empty).
-- NEVER label a chart or value "mock", "illustrative", "estimated", or "example" — if it is not real, do not render it. A widget must only ever show real, known values.
-- Educational math demos (e.g. compound-interest with user-given P/r/t) are fine because the user supplied the inputs; market data you don't have is NOT.
+# Data-source rule — strict (RAG on) vs. knowledge (RAG off). Selected in build_json_widget_rule().
+_DATA_RULE_STRICT = """
+NO FABRICATED DATA — ground every figure in the GROUNDING CONTEXT (applies to <RESPONSE> and <WIDGET>):
+- Use ONLY figures that appear in the GROUNDING CONTEXT provided below (retrieved financial data),
+  or numbers the user supplied in their message. Do NOT invent, estimate, recall from memory, or use
+  "illustrative"/"mock"/"example"/"est." data.
+- If the GROUNDING CONTEXT does not contain the figures needed to answer, say plainly in <RESPONSE>
+  that you do not have that data, suggest what to ask instead, and return <WIDGET></WIDGET> (empty).
+- NEVER label a chart or value "mock"/"illustrative"/"estimated" — if it is not in the context, do
+  not render it. A widget must only ever show real values from the GROUNDING CONTEXT.
+- Educational math demos (e.g. compound interest with user-given P/r/t) are fine — the user supplied
+  the inputs.
+"""
 
+_DATA_RULE_KNOWLEDGE = """
+DATA FROM YOUR KNOWLEDGE (applies to <RESPONSE> and <WIDGET>):
+- Use your own world knowledge to provide concrete, realistic figures and DRAW the widget. You do NOT
+  need a live source — give your best-known real-world values (e.g. company revenue, segments, market
+  caps) and render the chart rather than refusing.
+- If a number is approximate or from memory, you MAY briefly note that in <RESPONSE> (e.g. "approx",
+  "FY2023"), but still render the widget.
+- Do NOT fill charts with meaningless placeholder values (0,1,2,3 or random filler) — use genuine
+  best-estimate numbers that reflect reality.
+"""
+
+_JSON_RULE_TRAILER = """
 Data grounding (STRICT — the widget must mirror your <RESPONSE>):
 - Use ONLY the exact numbers and labels stated in your <RESPONSE>. Do NOT invent values, round differently, or add labels/values not written in <RESPONSE>.
 - For category charts, "x_categories" MUST be the exact entity/segment names you named in <RESPONSE> (e.g. "Data Center", "Gaming") — NEVER 0, 1, 2. Each series value MUST equal the number in <RESPONSE>, aligned to those categories.
@@ -436,7 +462,7 @@ Data grounding (STRICT — the widget must mirror your <RESPONSE>):
 
 Honoring an explicitly requested chart type:
 - If the user asks for a SPECIFIC chart type (e.g. "as a candlestick", "show a sankey", "pie chart") and it IS one of the supported kinds above, you MUST use exactly that kind — do not substitute another.
-- If the requested chart type is NOT in the supported kinds (e.g. 3D surface, map/choropleth, gantt, renko, marimekko, etc.), DO NOT silently render a different chart. Instead: in <RESPONSE>, say plainly that you cannot render that specific chart type yet, then name 1-3 supported kinds that would fit their data well and ask if they'd like one of those. In that case return <WIDGET></WIDGET> (empty) — wait for the user to confirm before rendering an alternative.
+- If the requested chart type is NOT in the supported kinds (e.g. 3D surface, map/choropleth, gantt, renko, marimekko, etc.), DO NOT silently render a different chart. Respond in a warm, professional, enterprise tone on behalf of the product "Vanguard" — briefly and politely explain that Vanguard's visualization suite does not currently support that chart type, then helpfully suggest 1-3 supported chart kinds that would present their data well and invite them to choose one. Return <WIDGET></WIDGET> (empty) and wait for them to pick. Do NOT use blunt or technical wording like "not supported in this system", "broken", "incorrect output", or "I cannot render" — keep it courteous and reassuring (e.g. "Vanguard doesn't support 3D surface charts just yet — but I'd be glad to show this as a 3D scatter, treemap, or heatmap instead. Which would you prefer?").
 - Never pretend an unsupported type is supported, and never relabel a different chart as the requested type.
 
 Interactivity:
@@ -446,9 +472,17 @@ Interactivity:
 - If something cannot be expressed with the supported block types (e.g. a playable game, live sliders), DO NOT invent HTML — return <WIDGET></WIDGET> (empty) and explain in <RESPONSE> instead. Never dump raw index arrays like `[0,1,2]`.
 - For photographs, diagrams, or icons, use the `image` block with a valid https:// URL or a data: URI. Combine `image` with `text`, `chart`, and `table` blocks as needed.
 
-Dynamic layout (JSON) — avoid static, single-block dashboards:
-- Shape `layout` like a short story: context first, then metrics, then detail, then actions. Mix block types whenever it improves scanning; do not default to one lonely chart if KPIs or a sentence of framing would help.
-- Use `action_row` for obvious follow-up intents; keep blocks ordered top-to-bottom by importance so the widget feels purposeful, not generic."""
+Dynamic layout (JSON) — informative, not padded:
+- ALWAYS provide proper DATA-ANALYST commentary with a chart. Put it as a `text` block in the widget BELOW the chart. The block's content MUST start with the heading line "### Analysis" then 3-4 NUMBERED points (each line: "1. ", "2. ", ...). Wrap the KEY figures/labels in **bold** (the renderer highlights bold in blue). Example content:
+"### Analysis
+1. **Data Center** at **$47.5B** is ~**78%** of total FY2024 revenue — the clear dominant segment.
+2. It grew ~**3.2x YoY** (from **$15.0B**), driven by AI/HPC demand.
+3. **Gaming** is a distant second at **$10.4B**; every other segment is under **$2B**.
+4. Takeaway: revenue is heavily **concentrated in AI infrastructure**."
+  Write like a financial/data analyst — specific and quantitative. Across the points cover: leader & laggard with values; the key gap/ratio/share/growth rate; the trend/notable change; any outlier or concentration; and a one-line "so what" takeaway. Use real numbers. NEVER write filler like "this chart shows the data" or "as you can see" — every point must carry a concrete insight. Always 3-4 points (not one). Never show a bare, unexplained chart.
+- Add a `table` when it genuinely HELPS the reader — e.g. exact reference figures, extra columns/notes, or a breakdown the chart can't show. Do NOT add a table that merely restates a simple chart's few values with nothing extra (that is clutter); use one when it adds real detail.
+- You MAY add a few KPI tiles for headline numbers when useful. Mix block types only when each adds value, and keep blocks ordered top-to-bottom by importance.
+- Use `action_row` for obvious follow-up intents so the widget feels purposeful, not generic."""
 
 
 def build_json_widget_rule() -> str:
@@ -473,7 +507,11 @@ def build_json_widget_rule() -> str:
             entries.append(f"{header}\n{body}")
         if not entries:
             return _JSON_WIDGET_RULE.strip()
-        return _JSON_RULE_PREAMBLE + "\n" + "\n".join(entries) + "\n" + _JSON_RULE_TRAILER
+        # Strict grounding when RAG is on; otherwise let the model use its own knowledge.
+        data_rule = _DATA_RULE_STRICT if getattr(config, "USE_RAG", False) else _DATA_RULE_KNOWLEDGE
+        return (
+            _JSON_RULE_PREAMBLE + "\n" + "\n".join(entries) + "\n" + data_rule + _JSON_RULE_TRAILER
+        )
     except Exception:
         return _JSON_WIDGET_RULE.strip()
 
@@ -577,6 +615,7 @@ def build_combined_system_prompt(
     format_rule: str,
     primitive_extra_context: str,
     user_message: str,
+    grounding_context: str = "",
     forbidden_components: list[str] | None = None,
     required_components: list[str] | None = None,
 ) -> str:
@@ -652,8 +691,30 @@ TOKEN LIMIT — you have ~{combined_max_tokens} tokens total for <RESPONSE> + <W
 - For comparison tables in <RESPONSE>: keep focused so the widget has room. Both must fit.
 """
 
+    if grounding_context.strip():
+        grounding_block = (
+            "\n═══════════════════════════════════════════════════════\n"
+            "GROUNDING CONTEXT — retrieved financial data (the ONLY source for your figures)\n"
+            "═══════════════════════════════════════════════════════\n"
+            "Use ONLY numbers/labels that appear below. Do not add figures from memory.\n\n"
+            f"{grounding_context.strip()}\n"
+        )
+    elif getattr(config, "USE_RAG", False):
+        # RAG on but nothing relevant retrieved → decline rather than fabricate.
+        grounding_block = (
+            "\n════════════════════��══════════════════════════════════\n"
+            "GROUNDING CONTEXT — none retrieved\n"
+            "═══════════════════════════════════════════════════════\n"
+            "No financial data was retrieved for this query. You do NOT have figures for it: "
+            "say so plainly in <RESPONSE> and return <WIDGET></WIDGET> (empty). Do NOT use memory or estimates.\n"
+        )
+    else:
+        # RAG off → knowledge mode (the data rule lets the model use its own figures).
+        grounding_block = ""
+
     return f"""You are an expert AI assistant. Each turn you produce a written answer AND an OPTIONAL interactive widget built ONLY from a fixed set of UI components defined below — there is NO HTML and NO external charting libraries.
 {social_turn_banner}
+{grounding_block}
 Output style: No emojis. Neat, clean, professional — in both <RESPONSE> text and <WIDGET>.
 {token_limit_block}
 {_OUTPUT_CONTRACT_STRICT}

vivek/backend/config.py

@@ -40,6 +40,17 @@ ADAPTIVE_LLM_MODE = os.getenv("ADAPTIVE_LLM_MODE", LLM_MODE).lower()
 # Widget output: "html" (full HTML in iframe) or "json" (UI schema for frontend renderer).
 WIDGET_MODE = os.getenv("WIDGET_MODE", "json").strip().lower()
 
+
+def _flag(name: str, default: str = "0") -> bool:
+    return os.getenv(name, default).strip().lower() in ("1", "true", "yes", "on")
+
+# Feature flags (default OFF):
+# - USE_BANDIT: when off, no Thompson-Sampling strategy selection or reward learning;
+#   the response uses a neutral style.
+# - USE_RAG: when off, no ChromaDB retrieval; the model fills figures from its own knowledge.
+USE_BANDIT = _flag("USE_BANDIT", "0")
+USE_RAG = _flag("USE_RAG", "0")
+
 # Max time (sec) and tokens for combined response+widget generation.
 # Tight defaults so latency stays predictable; raise only when the UI shows truncation.
 COMBINED_TIMEOUT_SECONDS = int(os.getenv("COMBINED_TIMEOUT_SECONDS", "45"))

vivek/backend/server.py

@@ -293,6 +293,19 @@ app.add_middleware(GZipMiddleware, minimum_size=1000)
 def _on_startup():
     persistence.init_db()
 
+    # RAG: only when enabled — ingest the financial corpus into ChromaDB on first boot
+    # (the binary store is not shipped). When USE_RAG is off, skip entirely (no model download).
+    if getattr(config, "USE_RAG", False):
+        try:
+            from rag_finance.retriever import _load as _rag_load
+            from rag_finance.ingest import ingest as _rag_ingest
+            col, _ = _rag_load()
+            if col is None or col.count() == 0:
+                stats = _rag_ingest()
+                print(f"[rag] ingested {stats.get('documents')} financial docs into ChromaDB")
+        except Exception as e:
+            print(f"[rag] ingest skipped: {e}")
+
     if getattr(config, "ADMIN_CONFIGURED", False) and config.ADMIN_USERNAME and config.ADMIN_PASSWORD and config.ADMIN_EMAIL:
         try:
             register_user(username=config.ADMIN_USERNAME, email=config.ADMIN_EMAIL, password=config.ADMIN_PASSWORD)
@@ -822,22 +835,29 @@ def _build_adaptive_prompt(uid: str, msg: str):
     ev = fast_valence(msg, user["last_response"])
     auto_detected = False
     auto_r = None
-
-    if user["last_response"] and user["last_x"] is not None and user["last_strategy"]:
-        reward = float(np.clip(0.5 + 0.45 * ev["pos"] - 0.45 * ev["neg"], 0.05, 0.95))
-        engine.update(uid, user["last_strategy"], np.array(user["last_x"]), reward)
-        auto_detected = True
-        auto_r = reward
-
     explicit = False
-    force_explore = bool(detect_explore_trigger(msg) or (ev.get("neg", 0.0) >= config.NEG_EXPLORE_THRESHOLD))
-    neg_s = negative_strength(ev)
-
-    strat, scores, x, prev = engine.select(
-        uid, msg, force_explore=force_explore, neg_strength=neg_s, explicit_strategy=None
-    )
-
-    format_rule = config.STRATEGIES.get(strat, "Be helpful and clear.")
+    force_explore = False
+
+    if config.USE_BANDIT:
+        # Implicit reward for the previous turn from this message's sentiment.
+        if user["last_response"] and user["last_x"] is not None and user["last_strategy"]:
+            reward = float(np.clip(0.5 + 0.45 * ev["pos"] - 0.45 * ev["neg"], 0.05, 0.95))
+            engine.update(uid, user["last_strategy"], np.array(user["last_x"]), reward)
+            auto_detected = True
+            auto_r = reward
+        force_explore = bool(detect_explore_trigger(msg) or (ev.get("neg", 0.0) >= config.NEG_EXPLORE_THRESHOLD))
+        neg_s = negative_strength(ev)
+        strat, scores, x, prev = engine.select(
+            uid, msg, force_explore=force_explore, neg_strength=neg_s, explicit_strategy=None
+        )
+        format_rule = config.STRATEGIES.get(strat, "Be helpful and clear.")
+    else:
+        # Bandit OFF: neutral style, no strategy selection, no learning.
+        strat = "neutral"
+        scores = {}
+        prev = None
+        x = engine.featurize(msg, user)  # still compute the feature vector for the Insights panel
+        format_rule = "Answer naturally, clearly, and concisely. There is no required text format."
 
     prim_block = ""
     if _is_admin_user(uid):
@@ -852,11 +872,24 @@ def _build_adaptive_prompt(uid: str, msg: str):
             if prim_lines:
                 prim_block = "\n\n## User primitives (follow these as constraints)\n" + "\n".join(prim_lines) + "\n"
 
+    # RAG: when ON, retrieve clean financial figures (entity-gated) → GROUNDING CONTEXT.
+    # When OFF, no context → the model fills figures from its own knowledge.
+    grounding_context = ""
+    if config.USE_RAG:
+        try:
+            from rag_finance.retriever import retrieve as _rag_retrieve
+            rag = _rag_retrieve(msg, k=6)
+            if rag.get("relevant") and rag.get("chunks"):
+                grounding_context = "\n\n".join(rag["chunks"])
+        except Exception as e:
+            print(f"[rag] retrieve failed: {e}")
+
     combined_system = build_combined_system_prompt(
         strategy_id=strat,
         format_rule=format_rule,
         primitive_extra_context=(getattr(config, "SKILLS_CONTENT", "") or "") + prim_block,
         user_message=msg,
+        grounding_context=grounding_context,
         forbidden_components=None,
         required_components=None,
     )

vivek/frontend-vue/dist/assets/{ActionRow-ByJHOpPb.js → ActionRow-yGTZNE6v.js}

@@ -1 +1 @@
-import{G as e,M as t,P as n,f as r,h as i,i as a,jt as o,u as s,v as c}from"./runtime-core.esm-bundler-olIhRSij.js";import{tr as l}from"./index-0cd4ouHe.js";var u={class:`flex flex-wrap gap-2`},d=c({__name:`ActionRow`,props:{block:{}},emits:[`action`],setup(c,{emit:d}){let f=d;function p(e){let t=String(e.label||e.intent||``).trim();t&&f(`action`,t)}return(d,f)=>(t(),r(`div`,u,[(t(!0),r(a,null,n(c.block.buttons||[],(n,r)=>(t(),s(l,{key:r,type:`button`,variant:`outline`,size:`sm`,title:n.intent?`Ask: ${n.label||n.intent}`:void 0,onClick:e=>p(n)},{default:e(()=>[i(o(n.label||`Action`),1)]),_:2},1032,[`title`,`onClick`]))),128))]))}});export{d as default};
+import{G as e,M as t,P as n,f as r,h as i,i as a,jt as o,u as s,v as c}from"./runtime-core.esm-bundler-olIhRSij.js";import{nr as l}from"./index-Dv2_3L6X.js";var u={class:`flex flex-wrap gap-2`},d=c({__name:`ActionRow`,props:{block:{}},emits:[`action`],setup(c,{emit:d}){let f=d;function p(e){let t=String(e.label||e.intent||``).trim();t&&f(`action`,t)}return(d,f)=>(t(),r(`div`,u,[(t(!0),r(a,null,n(c.block.buttons||[],(n,r)=>(t(),s(l,{key:r,type:`button`,variant:`outline`,size:`sm`,title:n.intent?`Ask: ${n.label||n.intent}`:void 0,onClick:e=>p(n)},{default:e(()=>[i(o(n.label||`Action`),1)]),_:2},1032,[`title`,`onClick`]))),128))]))}});export{d as default};

vivek/frontend-vue/dist/assets/TextBlock-BpmAPyAu.js

@@ -0,0 +1,3 @@
+import{M as e,c as t,f as n,v as r}from"./runtime-core.esm-bundler-olIhRSij.js";import{r as i}from"./index-Dv2_3L6X.js";var a=[`innerHTML`],o=i(r({__name:`TextBlock`,props:{block:{}},setup(r){let i=r;function o(e){return e.replace(/&/g,`&amp;`).replace(/</g,`&lt;`).replace(/>/g,`&gt;`)}function s(e){let t=o(e);return t=t.replace(/\*\*(.+?)\*\*/g,`<strong class="ws-accent">$1</strong>`),t=t.replace(/__(.+?)__/g,`<strong class="ws-accent">$1</strong>`),t=t.replace(/`([^`]+?)`/g,`<code class="ws-code">$1</code>`),t}let c=t(()=>{let e=String(i.block.content||``).replace(/\r\n/g,`
+`).trim();if(!e)return``;let t=e.split(`
+`),n=[],r=null,a=()=>{r&&=(n.push(`</${r}>`),null)};for(let e of t){let t=e.match(/^\s*#{1,3}\s+(.*)$/),i=e.match(/^\s*(\d+)[.)]\s+(.*)$/),o=e.match(/^\s*[-*•]\s+(.*)$/);t?(a(),n.push(`<div class="ws-h">${s(t[1])}</div>`)):i?(r!==`ol`&&(a(),n.push(`<ol class="ws-ol">`),r=`ol`),n.push(`<li>${s(i[2])}</li>`)):o?(r!==`ul`&&(a(),n.push(`<ul class="ws-bullets">`),r=`ul`),n.push(`<li>${s(o[1])}</li>`)):(a(),e.trim()&&n.push(`<p>${s(e)}</p>`))}return a(),n.join(``)});return(t,r)=>(e(),n(`div`,{class:`ws-text leading-relaxed motion-safe:transition-opacity motion-safe:duration-300`,innerHTML:c.value},null,8,a))}}),[[`__scopeId`,`data-v-51d1042a`]]);export{o as default};

vivek/frontend-vue/dist/assets/TextBlock-ChIy4qJQ.css

@@ -0,0 +1 @@
+.ws-text[data-v-51d1042a] p{margin:0 0 .45em}.ws-text[data-v-51d1042a] p:last-child{margin-bottom:0}.ws-text[data-v-51d1042a] .ws-h{letter-spacing:.04em;text-transform:uppercase;color:hsl(var(--muted-foreground));margin:.2em 0 .5em;font-size:.8rem;font-weight:700}.ws-text[data-v-51d1042a] .ws-bullets{flex-direction:column;gap:.4em;margin:.25em 0;padding-left:0;list-style:none;display:flex}.ws-text[data-v-51d1042a] .ws-bullets li{padding-left:1.1em;position:relative}.ws-text[data-v-51d1042a] .ws-bullets li:before{content:"";background:hsl(var(--primary));border-radius:9999px;width:6px;height:6px;position:absolute;top:.55em;left:0}.ws-text[data-v-51d1042a] .ws-ol{counter-reset:ws;flex-direction:column;gap:.5em;margin:.25em 0;padding-left:0;list-style:none;display:flex}.ws-text[data-v-51d1042a] .ws-ol li{counter-increment:ws;padding-left:1.9em;position:relative}.ws-text[data-v-51d1042a] .ws-ol li:before{content:counter(ws);width:1.35em;height:1.35em;color:hsl(var(--primary));background:hsl(var(--primary) / .12);border-radius:9999px;justify-content:center;align-items:center;font-size:.72em;font-weight:700;display:inline-flex;position:absolute;top:.05em;left:0}.ws-text[data-v-51d1042a] .ws-accent{color:hsl(var(--primary));font-weight:600}.ws-text[data-v-51d1042a] .ws-code{background:hsl(var(--primary) / .1);color:hsl(var(--primary));border-radius:.3rem;padding:.05em .35em;font-size:.88em}

vivek/frontend-vue/dist/assets/TextBlock-CoLO1TqP.js

@@ -1 +0,0 @@
-import{M as e,f as t,jt as n,v as r}from"./runtime-core.esm-bundler-olIhRSij.js";var i={class:`whitespace-pre-wrap leading-relaxed motion-safe:transition-opacity motion-safe:duration-300`},a=r({__name:`TextBlock`,props:{block:{}},setup(r){return(a,o)=>(e(),t(`div`,i,n(r.block.content||``),1))}});export{a as default};

vivek/frontend-vue/dist/index.html

@@ -5,9 +5,9 @@
     <link rel="icon" type="image/svg+xml" href="/favicon.svg" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>frontend-vue</title>
-    <script type="module" crossorigin src="/assets/index-0cd4ouHe.js"></script>
+    <script type="module" crossorigin src="/assets/index-Dv2_3L6X.js"></script>
     <link rel="modulepreload" crossorigin href="/assets/runtime-core.esm-bundler-olIhRSij.js">
-    <link rel="stylesheet" crossorigin href="/assets/index-yy7aehFb.css">
+    <link rel="stylesheet" crossorigin href="/assets/index-CKwtM7do.css">
   </head>
   <body>
     <div id="app"></div>

vivek/frontend-vue/src/components/WidgetSchemaChart.vue

@@ -96,7 +96,7 @@ watch(() => props.chart, refresh, { deep: true })
         <ArrowDownTrayIcon class="h-3 w-3" /> PNG
       </button>
     </div>
-    <div v-if="renderable" ref="rootEl" class="w-full h-[min(360px,52vh)] min-h-[220px]" />
+    <div v-if="renderable" ref="rootEl" class="w-full h-[clamp(340px,56vh,560px)] min-h-[300px]" />
     <div v-else class="px-3 py-6 text-xs text-muted-foreground text-center">
       No renderable data for chart type "{{ kind }}".
     </div>

vivek/frontend-vue/src/components/widgets/TextBlock.vue

@@ -1,9 +1,150 @@
 <script setup lang="ts">
-defineProps<{ block: { content?: string } }>()
+import { computed } from 'vue'
+
+const props = defineProps<{ block: { content?: string } }>()
+
+// Escape first (safe), then apply a tiny markdown subset so the model's analysis can use
+// "- " bullets and **bold**/`$ figures` highlighted in the accent color. No raw HTML allowed.
+function esc(s: string): string {
+  return s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+}
+
+function inline(s: string): string {
+  let t = esc(s)
+  // **bold** and __bold__ → accent (blue) emphasis
+  t = t.replace(/\*\*(.+?)\*\*/g, '<strong class="ws-accent">$1</strong>')
+  t = t.replace(/__(.+?)__/g, '<strong class="ws-accent">$1</strong>')
+  // `code` / inline figures
+  t = t.replace(/`([^`]+?)`/g, '<code class="ws-code">$1</code>')
+  return t
+}
+
+const html = computed(() => {
+  const raw = String(props.block.content || '').replace(/\r\n/g, '\n').trim()
+  if (!raw) return ''
+  const lines = raw.split('\n')
+  const out: string[] = []
+  let list: 'ul' | 'ol' | null = null
+  const closeList = () => {
+    if (list) {
+      out.push(`</${list}>`)
+      list = null
+    }
+  }
+  for (const line of lines) {
+    const heading = line.match(/^\s*#{1,3}\s+(.*)$/)
+    const ordered = line.match(/^\s*(\d+)[.)]\s+(.*)$/)
+    const bullet = line.match(/^\s*[-*•]\s+(.*)$/)
+    if (heading) {
+      closeList()
+      out.push(`<div class="ws-h">${inline(heading[1])}</div>`)
+    } else if (ordered) {
+      if (list !== 'ol') {
+        closeList()
+        out.push('<ol class="ws-ol">')
+        list = 'ol'
+      }
+      out.push(`<li>${inline(ordered[2])}</li>`)
+    } else if (bullet) {
+      if (list !== 'ul') {
+        closeList()
+        out.push('<ul class="ws-bullets">')
+        list = 'ul'
+      }
+      out.push(`<li>${inline(bullet[1])}</li>`)
+    } else {
+      closeList()
+      if (line.trim()) out.push(`<p>${inline(line)}</p>`)
+    }
+  }
+  closeList()
+  return out.join('')
+})
 </script>
 
 <template>
-  <div class="whitespace-pre-wrap leading-relaxed motion-safe:transition-opacity motion-safe:duration-300">
-    {{ block.content || '' }}
-  </div>
+  <div
+    class="ws-text leading-relaxed motion-safe:transition-opacity motion-safe:duration-300"
+    v-html="html"
+  />
 </template>
+
+<style scoped>
+.ws-text :deep(p) {
+  margin: 0 0 0.45em;
+}
+.ws-text :deep(p:last-child) {
+  margin-bottom: 0;
+}
+.ws-text :deep(.ws-h) {
+  font-size: 0.8rem;
+  font-weight: 700;
+  letter-spacing: 0.04em;
+  text-transform: uppercase;
+  color: hsl(var(--muted-foreground));
+  margin: 0.2em 0 0.5em;
+}
+.ws-text :deep(.ws-bullets) {
+  margin: 0.25em 0;
+  padding-left: 0;
+  list-style: none;
+  display: flex;
+  flex-direction: column;
+  gap: 0.4em;
+}
+.ws-text :deep(.ws-bullets li) {
+  position: relative;
+  padding-left: 1.1em;
+}
+.ws-text :deep(.ws-bullets li)::before {
+  content: '';
+  position: absolute;
+  left: 0;
+  top: 0.55em;
+  width: 6px;
+  height: 6px;
+  border-radius: 9999px;
+  background: hsl(var(--primary));
+}
+.ws-text :deep(.ws-ol) {
+  margin: 0.25em 0;
+  padding-left: 0;
+  list-style: none;
+  counter-reset: ws;
+  display: flex;
+  flex-direction: column;
+  gap: 0.5em;
+}
+.ws-text :deep(.ws-ol li) {
+  position: relative;
+  padding-left: 1.9em;
+  counter-increment: ws;
+}
+.ws-text :deep(.ws-ol li)::before {
+  content: counter(ws);
+  position: absolute;
+  left: 0;
+  top: 0.05em;
+  width: 1.35em;
+  height: 1.35em;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  font-size: 0.72em;
+  font-weight: 700;
+  border-radius: 9999px;
+  color: hsl(var(--primary));
+  background: hsl(var(--primary) / 0.12);
+}
+.ws-text :deep(.ws-accent) {
+  color: hsl(var(--primary));
+  font-weight: 600;
+}
+.ws-text :deep(.ws-code) {
+  font-size: 0.88em;
+  padding: 0.05em 0.35em;
+  border-radius: 0.3rem;
+  background: hsl(var(--primary) / 0.1);
+  color: hsl(var(--primary));
+}
+</style>

vivek/frontend-vue/src/lib/echartsOption.ts

@@ -706,8 +706,11 @@ export function buildEChartsOption(chart: ChartSpec, title = '', opts: { dark?:
         type: 'category' as const,
         data: cats,
         name: (isHBar ? chart?.y_label : chart?.x_label) || '',
-        nameLocation: 'middle' as const,
-        nameGap: 30,
+        // For hbar the category axis is vertical, so a centered axis name overlaps the
+        // (often long) category labels — place it at the axis end instead.
+        nameLocation: (isHBar ? 'end' : 'middle') as 'end' | 'middle',
+        nameGap: isHBar ? 12 : (cats as string[]).length > 6 ? 42 : 28,
+        nameTextStyle: isHBar ? { align: 'right' as const, color: tc } : { color: tc },
         axisLabel: { fontSize: 10, rotate: !isHBar && (cats as string[]).length > 6 ? 30 : 0 },
       }
     : {

vivek/frontend-vue/src/lib/exportWidgetHtml.ts

@@ -176,8 +176,8 @@ function renderBlock(block: Block, idx: number, charts: ChartEntry[]): string {
 
 const STYLES = `
 *{box-sizing:border-box}
-body{margin:0;background:#f7f8fa;color:#111318;font:14px/1.5 system-ui,-apple-system,Segoe UI,Roboto,sans-serif;padding:24px}
-.ws-doc{max-width:880px;margin:0 auto;display:flex;flex-direction:column;gap:14px}
+body{margin:0;background:#f7f8fa;color:#111318;font:15px/1.6 system-ui,-apple-system,Segoe UI,Roboto,sans-serif;padding:clamp(16px,3vw,40px)}
+.ws-doc{max-width:1200px;width:96vw;margin:0 auto;display:flex;flex-direction:column;gap:18px}
 .ws-card{background:#fff;border:1px solid #e6e8ec;border-radius:14px;padding:16px;box-shadow:0 1px 2px rgba(0,0,0,.04)}
 .ws-text{margin:0;color:#2b2f38}
 .ws-h{font-weight:600;font-size:13px;margin-bottom:8px}
@@ -202,10 +202,10 @@ body{margin:0;background:#f7f8fa;color:#111318;font:14px/1.5 system-ui,-apple-sy
 .ws-action-row{display:flex;flex-wrap:wrap;gap:8px}
 .ws-action{font-size:13px;padding:6px 12px;border:1px solid #d7dae0;border-radius:8px;background:#fff;color:#475569}
 .ws-fig{margin:0}
-.ws-fig img{width:100%;max-height:420px;border-radius:10px}
+.ws-fig img{width:100%;max-height:min(70vh,640px);object-fit:contain;border-radius:10px}
 .ws-fig figcaption{font-size:12px;color:#9aa1ad;margin-top:6px;text-align:center}
-.ws-chart{width:100%;height:380px}
-.ws-chart-3d{height:460px}
+.ws-chart{width:100%;height:clamp(420px,68vh,720px)}
+.ws-chart-3d{height:clamp(480px,78vh,820px)}
 .ws-note{font-size:12px;color:#9aa1ad}
 .ws-footer{text-align:center;font-size:11px;color:#b3b9c4;margin-top:6px}
 `

vivek/frontend-vue/src/widget-registry.json

@@ -3,9 +3,9 @@
   "blocks": [
     {
       "type": "text",
-      "whenToUse": "Narrative explanation, framing sentence, or any prose paragraph.",
+      "whenToUse": "Narrative explanation, analysis, or framing. Supports a light markdown subset in content: \"- \" bullet lines and **bold** (bold renders highlighted in the accent/blue color). Great for analyst commentary as bullet points.",
       "spec": [
-        "{ \"type\": \"text\", \"id\": \"...\", \"content\": \"...\" }"
+        "{ \"type\": \"text\", \"id\": \"...\", \"content\": \"- **Data Center** at **$47.5B** is ~**78%** of total\\n- **Gaming** recovered to **$10.4B** (+14% YoY)\" }"
       ],
       "example": {
         "type": "text",

vivek/rag_finance/retriever.py

@@ -87,27 +87,32 @@ def retrieve(query: str, k: int = 6) -> dict:
             return _empty(query)
     except Exception:
         pass
-    # Retrieve a wider pool so entity filtering can find the right company's docs.
-    res = col.query(query_texts=[_clean_for_retrieval(query)], n_results=max(k, 12))
-    ids = res["ids"][0]
-    docs = res["documents"][0]
-    metas = res["metadatas"][0]
-    dists = res["distances"][0]
-    rows = list(zip(ids, docs, metas, dists))
-
-    # ENTITY-AWARE RELEVANCE GATE:
-    # - Query names a company we HAVE → keep that company's docs (let the synthesizer
-    #   handle any missing metric). Distance alone is unreliable for minor metrics.
-    # - Query names NO known company → distance gate (catches off-topic / unknown companies).
+    # ENTITY-AWARE RETRIEVAL:
+    # - Query NAMES companies we have → fetch ALL their docs (every fiscal year) by metadata,
+    #   so multi-company comparisons get full coverage (embedding top-k can miss some).
+    #   The synthesizer then picks the right year(s).
+    # - Query names NO known company → embedding search + distance gate (catches off-topic).
     mentioned = _mentioned_ids(query)
     if mentioned:
-        kept = [r for r in rows if r[2].get("company_id") in mentioned]
-        if not kept:  # company named but its docs weren't in the pool → loose distance fallback
-            kept = [r for r in rows if r[3] <= ENTITY_THRESHOLD]
+        try:
+            got = col.get(where={"company_id": {"$in": list(mentioned)}})
+            gids = got.get("ids") or []
+            gdocs = got.get("documents") or []
+            gmetas = got.get("metadatas") or []
+            kept = list(zip(gids, gdocs, gmetas, [0.0] * len(gids)))
+        except Exception:
+            kept = []
+        if not kept:  # fallback: embedding pool filtered to mentioned companies
+            res = col.query(query_texts=[_clean_for_retrieval(query)], n_results=24)
+            rows = list(zip(res["ids"][0], res["documents"][0], res["metadatas"][0], res["distances"][0]))
+            kept = [r for r in rows if r[2].get("company_id") in mentioned] or [
+                r for r in rows if r[3] <= ENTITY_THRESHOLD
+            ]
+        kept = kept[:21]  # all docs for up to ~7 companies × 3 years
     else:
-        kept = [r for r in rows if r[3] <= RELEVANCE_THRESHOLD]
-
-    kept = kept[:9]  # cap context size
+        res = col.query(query_texts=[_clean_for_retrieval(query)], n_results=max(k, 12))
+        rows = list(zip(res["ids"][0], res["documents"][0], res["metadatas"][0], res["distances"][0]))
+        kept = [r for r in rows if r[3] <= RELEVANCE_THRESHOLD][:9]
     relevant = len(kept) > 0
 
     doc_ids = [x[0] for x in kept]
@@ -124,7 +129,7 @@ def retrieve(query: str, k: int = 6) -> dict:
         "metas": kept_metas,
         "evidence": packet,
         "relevant": relevant,
-        "best_distance": min(dists) if dists else None,
+        "best_distance": (min((x[3] for x in kept), default=None) if kept else None),
     }
 
 

vivek/requirements.txt

@@ -6,3 +6,4 @@ uvicorn[standard]
 SQLAlchemy
 PyJWT
 bcrypt
+chromadb>=1.0