feat(briefing): cache + map-reduce article context for rich discuss chats

The Discuss button on news cards was producing one-shot replies because
the model got the whole trafilatura blob dropped into history with a
canned "summarize and discuss this article" prompt — no length guard, no
prep, no invitation to converse. Large articles got silently truncated by
Ollama; small articles got a tepid reply.

This reworks discuss_article around a three-layer cache:

  context_prepared  →  content_full  →  fresh trafilatura fetch

First click on a small article fetches once, writes through to both
caches, and passes the body straight into the synthetic read_article
tool-result. First click on a large article additionally runs a parallel
map step (services/article_context.py) that chunks the body on paragraph
boundaries, summarizes each ~8k chunk to ~300 words of dense factual
prose via the background model, and concatenates the summaries under
section headers — all pinned to num_ctx=16384 so the map step doesn't
itself fall victim to silent truncation. Repeat clicks on either path
skip straight to the chat turn.

The canned summary prompt is replaced with a conversational seed that
invites the user into an actual discussion rather than a one-shot
synopsis, matching the goal of "have a conversation about an article,
not just read it."

discuss_topic is intentionally left untouched — it's the multi-article
aggregation path and needs a separate rework. Follow-up task will decide
whether to retire it or rework it on the cached-context approach.

Closes task #106.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

alembic/versions/0038_add_rss_item_full_content.py

@@ -0,0 +1,34 @@
+"""Add content_full and context_prepared caches to rss_items.
+
+Revision ID: 0038
+Revises: 0037
+Create Date: 2026-04-13
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0038"
+down_revision = "0037"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("rss_items", sa.Column("content_full", sa.Text(), nullable=True))
+    op.add_column(
+        "rss_items",
+        sa.Column("context_prepared", sa.Text(), nullable=True),
+    )
+    op.add_column(
+        "rss_items",
+        sa.Column("content_fetched_at", sa.DateTime(timezone=True), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("rss_items", "content_fetched_at")
+    op.drop_column("rss_items", "context_prepared")
+    op.drop_column("rss_items", "content_full")

src/fabledassistant/models/rss_feed.py

@@ -46,6 +46,17 @@ class RssItem(Base):
     published_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
     # Truncated to 2000 chars to keep DB size reasonable
     content: Mapped[str] = mapped_column(Text, default="")
+    # Full trafilatura-extracted article body, populated lazily on first
+    # discuss-click / enrichment pass. Nullable — most items never get this
+    # cached. Expires naturally with the item (90-day retention).
+    content_full: Mapped[str | None] = mapped_column(Text, nullable=True)
+    # Map-reduced conversation-ready context derived from content_full. See
+    # services/article_context.py — populated on first discuss click so
+    # repeat clicks skip both the fetch and the LLM map step.
+    context_prepared: Mapped[str | None] = mapped_column(Text, nullable=True)
+    content_fetched_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True
+    )
     fetched_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), default=lambda: datetime.now(timezone.utc)
     )

src/fabledassistant/routes/briefing.py

@@ -532,8 +532,28 @@ async def discuss_article(item_id: int):
     if get_buffer(conv_id) is not None:
         return jsonify({"error": "Generation already in progress"}), 409
 
-    from fabledassistant.services.rss import _fetch_full_article
-    article_content = await _fetch_full_article(item.url) or item.content or ""
+    # Three-layer cache: context_prepared (post-map-reduce) → content_full
+    # (raw trafilatura) → fresh fetch. Only the first miss pays the fetch
+    # cost; only a large uncached article pays the map-reduce cost. Repeat
+    # clicks on the same article skip straight to the chat turn.
+    from fabledassistant.services.article_context import prepare_article_context
+    from fabledassistant.services.rss import get_or_fetch_full_article
+
+    model = await get_setting(uid, "default_model", "") or ""
+
+    if item.context_prepared:
+        article_content = item.context_prepared
+    else:
+        raw_body = await get_or_fetch_full_article(item) or item.content or ""
+        article_content = await prepare_article_context(
+            item.title or "", item.url, raw_body, model,
+        )
+        if article_content:
+            async with async_session() as session:
+                fresh = await session.get(RssItem, item.id)
+                if fresh is not None:
+                    fresh.context_prepared = article_content
+                    await session.commit()
 
     # Store synthetic assistant message with read_article tool result
     synthetic_tool_calls = [{
@@ -549,8 +569,17 @@ async def discuss_article(item_id: int):
     }]
     await add_message(conv_id, "assistant", "", status="complete", tool_calls=synthetic_tool_calls)
 
-    # Store user message
-    await add_message(conv_id, "user", "Please summarize and discuss this article.")
+    # Conversational seed — invites a real discussion rather than asking for
+    # a one-shot summary. The model sees the article context in the tool
+    # result above and responds to this user turn as the start of an ongoing
+    # conversation the user will steer with follow-ups.
+    discuss_prompt = (
+        "I want to talk about this article. Start with a substantive summary "
+        "of what it's arguing and the key evidence it uses, then tell me what "
+        "stood out to you or seems worth pushing back on. I'll ask follow-ups "
+        "from there."
+    )
+    await add_message(conv_id, "user", discuss_prompt)
 
     # Reload conversation with fresh messages to build history
     conv = await get_conversation(uid, conv_id)
@@ -572,15 +601,13 @@ async def discuss_article(item_id: int):
         else:
             history.append(msg_dict)
 
-    model = await get_setting(uid, "default_model", "") or ""
-
     assistant_msg = await add_message(conv_id, "assistant", "", status="generating")
     buf = create_buffer(conv_id, assistant_msg.id)
 
     asyncio.create_task(run_generation(
         buf, history, model,
         uid, conv_id, conv.title or "",
-        "Please summarize and discuss this article.",
+        discuss_prompt,
     ))
 
     return jsonify({"assistant_message_id": assistant_msg.id, "status": "generating"}), 202

src/fabledassistant/services/article_context.py

@@ -0,0 +1,161 @@
+"""Prepare article bodies as conversation-ready context.
+
+Used by the briefing ``discuss-article`` flow. A raw trafilatura extraction
+is often too large to drop whole into a chat history without eating the
+context window, so this module runs a map-reduce step over oversized
+articles and returns a compact, structured context that still preserves the
+article's meaning across sections.
+
+Small articles pass through unchanged — map-reduce only fires when the raw
+body exceeds CHAR_BUDGET. The output is cached on ``rss_items.context_prepared``
+by the caller, so repeat discuss-clicks on the same article skip this work
+entirely.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import re
+
+from fabledassistant.services.llm import generate_completion
+
+logger = logging.getLogger(__name__)
+
+# ~12k tokens at 4 chars/token. Comfortably under OLLAMA_NUM_CTX=16384
+# with room left for system prompt, chat history, and the assistant reply.
+CHAR_BUDGET = 48_000
+
+# Chunk size for the map step on oversized articles. Overlap preserves
+# context across paragraph boundaries that happen to land mid-sentence.
+CHUNK_CHARS = 8_000
+CHUNK_OVERLAP = 400
+
+_PARA_SPLIT = re.compile(r"\n\s*\n")
+
+
+def _chunk_by_paragraph(body: str) -> list[str]:
+    """Split ``body`` into chunks of up to CHUNK_CHARS, respecting paragraphs.
+
+    Paragraphs longer than CHUNK_CHARS are split mid-paragraph as a last
+    resort. Adjacent chunks share CHUNK_OVERLAP chars of trailing text so
+    a sentence straddling the boundary stays readable on both sides.
+    """
+    paragraphs = [p.strip() for p in _PARA_SPLIT.split(body) if p.strip()]
+    chunks: list[str] = []
+    current: list[str] = []
+    current_len = 0
+    for para in paragraphs:
+        para_len = len(para)
+        if para_len > CHUNK_CHARS:
+            if current:
+                chunks.append("\n\n".join(current))
+                current, current_len = [], 0
+            for i in range(0, para_len, CHUNK_CHARS - CHUNK_OVERLAP):
+                chunks.append(para[i : i + CHUNK_CHARS])
+            continue
+        if current_len + para_len + 2 > CHUNK_CHARS and current:
+            chunks.append("\n\n".join(current))
+            tail = current[-1][-CHUNK_OVERLAP:] if current else ""
+            current = [tail, para] if tail else [para]
+            current_len = len(tail) + para_len + (2 if tail else 0)
+        else:
+            current.append(para)
+            current_len += para_len + 2
+    if current:
+        chunks.append("\n\n".join(current))
+    return chunks
+
+
+async def _summarize_chunk(title: str, chunk: str, index: int, total: int, model: str) -> str:
+    """Map-step summary of one article chunk.
+
+    Aims for ~300 words of dense, factual prose — not bullet points — so the
+    downstream chat model can quote from it naturally.
+    """
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "You are summarizing one section of a larger article so a downstream "
+                "conversation model can discuss the full article without having to read "
+                "every word.\n\n"
+                "Requirements:\n"
+                "- 250–350 words of dense factual prose\n"
+                "- Preserve specific claims, numbers, names, and quotes\n"
+                "- Do NOT editorialize or add analysis\n"
+                "- Do NOT use bullet points or headings\n"
+                "- Do NOT say 'this section' or 'this article' — write content, not meta"
+            ),
+        },
+        {
+            "role": "user",
+            "content": (
+                f"Article: {title}\n"
+                f"Section {index + 1} of {total}:\n\n{chunk}"
+            ),
+        },
+    ]
+    try:
+        # Pin num_ctx — same rationale as services/research.py:66. A large
+        # chunk plus system prompt can push well past the default window;
+        # silent truncation here would drop the tail of the chunk without
+        # any error, producing a misleading summary.
+        raw = await generate_completion(
+            messages, model, max_tokens=600, num_ctx=16384
+        )
+        return raw.strip()
+    except Exception:
+        logger.warning(
+            "Article chunk summary failed for section %d/%d of '%s'",
+            index + 1, total, title, exc_info=True,
+        )
+        # Fall back to the raw chunk truncated to ~1500 chars so the overall
+        # pipeline still delivers something rather than dropping the section.
+        return chunk[:1500]
+
+
+async def prepare_article_context(
+    title: str,
+    url: str,
+    body: str,
+    model: str,
+) -> str:
+    """Return a conversation-ready context block for ``body``.
+
+    - Small article (≤ CHAR_BUDGET): returns ``body`` unchanged.
+    - Oversized article: runs a parallel map step over paragraph-aware
+      chunks and concatenates the summaries under section headers.
+
+    The returned string is what should go into the ``read_article`` synthetic
+    tool-result in chat history. Callers are responsible for caching it to
+    ``rss_items.context_prepared``.
+    """
+    body = body or ""
+    if len(body) <= CHAR_BUDGET:
+        return body
+
+    chunks = _chunk_by_paragraph(body)
+    logger.info(
+        "Article '%s' is %d chars, map-reducing into %d chunks",
+        title, len(body), len(chunks),
+    )
+
+    summaries = await asyncio.gather(
+        *[
+            _summarize_chunk(title, chunk, i, len(chunks), model)
+            for i, chunk in enumerate(chunks)
+        ]
+    )
+
+    header = (
+        f"(This article was longer than the chat window could hold verbatim, "
+        f"so the full text was split into {len(chunks)} sections and each was "
+        "summarized below. Each section preserves specific claims, numbers, "
+        "and quotes from the original.)\n\n"
+    )
+    parts = [
+        f"## Section {i + 1}\n\n{summary}"
+        for i, summary in enumerate(summaries)
+    ]
+    return header + "\n\n".join(parts)

src/fabledassistant/services/rss.py

@@ -34,6 +34,34 @@ def _html_to_text(html: str) -> str:
         return html
 
 
+async def get_or_fetch_full_article(item: RssItem) -> str | None:
+    """Return the full article body, fetching+caching on miss.
+
+    Checks ``item.content_full`` first — populated either by the enrichment
+    pass at feed-ingest time or by a previous discuss-click. On miss, fetches
+    via ``_fetch_full_article`` and writes through. Returns ``None`` only if
+    the fetch itself fails; ``item.content_full == ""`` is still a cache hit.
+
+    Callers must pass an RssItem attached to an open session if they want
+    the write-through to persist — otherwise the fetched text is returned
+    but the cache stays empty and the next click will re-fetch.
+    """
+    if item.content_full is not None:
+        return item.content_full
+    if not item.url:
+        return None
+    text = await _fetch_full_article(item.url)
+    if text is None:
+        return None
+    async with async_session() as session:
+        fresh = await session.get(RssItem, item.id)
+        if fresh is not None:
+            fresh.content_full = text
+            fresh.content_fetched_at = datetime.now(timezone.utc)
+            await session.commit()
+    return text
+
+
 async def _fetch_full_article(url: str) -> str | None:
     """Fetch a URL and extract its main article text via trafilatura.
 
@@ -209,6 +237,11 @@ async def fetch_and_cache_feed(feed_id: int, url: str) -> int:
                         item = await session.get(RssItem, item_id)
                         if item:
                             item.content = full_text
+                            # Populate the discuss-click cache too so the
+                            # first click skips straight to the map-reduce
+                            # step without re-fetching.
+                            item.content_full = full_text
+                            item.content_fetched_at = datetime.now(timezone.utc)
                             await session.commit()
                             await upsert_rss_item_embedding(
                                 item_id, feed_user_id, item.title or "", item.content

tests/test_article_reading.py

@@ -134,3 +134,83 @@ def test_history_builder_no_tool_calls_unchanged():
     assert len(history) == 2
     assert history[0] == {"role": "user", "content": "Hello"}
     assert history[1] == {"role": "assistant", "content": "Hi there!"}
+
+
+# ---------------------------------------------------------------------------
+# prepare_article_context tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_prepare_article_context_small_passthrough():
+    """Articles under CHAR_BUDGET pass through unchanged with zero LLM calls."""
+    from fabledassistant.services import article_context
+
+    body = "A short article.\n\nWith two paragraphs."
+    with patch(
+        "fabledassistant.services.article_context.generate_completion",
+        new_callable=AsyncMock,
+    ) as mock_gen:
+        out = await article_context.prepare_article_context(
+            "Title", "https://example.com", body, "test-model",
+        )
+
+    assert out == body
+    mock_gen.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_prepare_article_context_large_runs_map_reduce():
+    """Articles over CHAR_BUDGET are chunked and map-reduced via the background model."""
+    from fabledassistant.services import article_context
+
+    # CHAR_BUDGET is 48_000 — build a body well over that with paragraph breaks
+    # so the chunker has natural splits to work with.
+    paragraph = ("Lorem ipsum dolor sit amet, consectetur adipiscing elit. " * 40).strip()
+    body = "\n\n".join([paragraph] * 30)  # ~70k+ chars across 30 paragraphs
+    assert len(body) > article_context.CHAR_BUDGET
+
+    with patch(
+        "fabledassistant.services.article_context.generate_completion",
+        new_callable=AsyncMock,
+        return_value="Summary of this section with specific claims preserved.",
+    ) as mock_gen:
+        out = await article_context.prepare_article_context(
+            "Long Article", "https://example.com/long", body, "test-model",
+        )
+
+    # At least one LLM call fired (the map step ran)
+    assert mock_gen.await_count >= 1
+    # Output carries the oversized-article header and section markers
+    assert "longer than the chat window" in out
+    assert "## Section 1" in out
+    # Map output is much smaller than the raw body
+    assert len(out) < len(body)
+
+
+def test_chunk_by_paragraph_respects_boundaries():
+    """Chunker splits on paragraph breaks, not mid-sentence."""
+    from fabledassistant.services.article_context import _chunk_by_paragraph, CHUNK_CHARS
+
+    paragraphs = [f"Paragraph {i}. " + ("x" * 1000) for i in range(20)]
+    body = "\n\n".join(paragraphs)
+
+    chunks = _chunk_by_paragraph(body)
+
+    # Each chunk stays under the budget
+    for chunk in chunks:
+        assert len(chunk) <= CHUNK_CHARS
+    # Total content is preserved (modulo overlap duplication, so ≥ original)
+    assert sum(len(c) for c in chunks) >= len(body) * 0.9
+
+
+def test_chunk_by_paragraph_handles_oversized_paragraph():
+    """A single paragraph larger than CHUNK_CHARS gets split mid-paragraph."""
+    from fabledassistant.services.article_context import _chunk_by_paragraph, CHUNK_CHARS
+
+    body = "x" * (CHUNK_CHARS * 3)  # one huge paragraph, no breaks
+    chunks = _chunk_by_paragraph(body)
+
+    assert len(chunks) >= 3
+    for chunk in chunks:
+        assert len(chunk) <= CHUNK_CHARS