A user has 47 sales calls recorded over the past quarter. They are organized in a folder called "Q1 Enterprise Pipeline." They want to know: "Which prospects mentioned budget constraints, and what was the common objection pattern?"
Before folder intelligence, the answer required opening each conversation, reading the summary, and manually synthesizing across all 47. That is not reasonable. No one does it. The information exists, but it is siloed inside individual conversations.
Now they type the question, and the system searches across all 47 conversations, finds the relevant passages, and produces an answer with citations. Each citation links to a specific conversation, timestamp, and speaker.
This is the feature that transforms Neural Summary from a document generator into a knowledge base.
The problem: information silos
Every conversation in Neural Summary produces a rich set of structured data: a summary with key points, detailed sections, decisions, next steps, and any generated lenses (action items, CRM notes, executive summaries, etc.).
But this data lives inside individual conversations. If you want to find a pattern that spans multiple conversations, you have to look at each one manually. The data is structured, searchable, and indexed within a single conversation. Across conversations, it is invisible.
This is the same problem that exists in every document-based knowledge system. Each document is well-organized internally. The cross-document view is absent.
How it works
Folder intelligence has three components: vector indexing, cross-conversation search, and streaming chat with citations.
Vector indexing. When a conversation is processed, we index its full summary content into a vector database. Not just the title and headline, but key point descriptions, detailed section content, decisions, next steps, and any generated lens output. Each indexed chunk carries metadata: the conversation ID, the conversation title, and the source field.
When a user first asks a question in a folder, any conversations that have not been indexed yet are indexed on the fly. A status message shows progress: "Reading through 47 conversations, this can take a few seconds."
Cross-conversation search. When the user asks a question, we convert it to a vector embedding and search across all indexed conversations in the folder. The search returns the most relevant passages, each tagged with its source conversation.
These passages become the context for the LLM. The prompt includes the question, the relevant passages with their conversation metadata, and instructions to produce a synthesized answer with citations.
Streaming chat. The response streams token by token using Server-Sent Events. As the LLM generates its answer, it includes citations in the format [Conversation Title — MM:SS, Speaker Name]. The frontend renders these as interactive badges that link to the specific moment in the source conversation.
The citation system
Citations are the feature that makes folder intelligence trustworthy. Without citations, a synthesized answer across 47 conversations is an unverifiable claim. With citations, every statement is traceable to a specific source.
The citation format differs between conversation-level and folder-level chat:
Conversation chat: [12:45, Sarah Johnson]
Folder chat: [Q1 Pipeline Review — 12:45, Sarah Johnson]
The folder citation includes the conversation title because the user needs to know which of the 47 conversations the reference comes from.
In the frontend, citations render as styled badges. Clicking a conversation-level citation jumps to that timestamp in the transcript. Clicking a folder-level citation navigates to the source conversation at the referenced timestamp.
One subtle challenge: citation buffering during streaming. The LLM streams its response token by token. A citation like [Q1 Pipeline Review — 12:45, Sarah] arrives across multiple tokens. If we render tokens immediately, the user would see fragments like [Q1 Pipel mid-stream, which looks broken.
Our citation buffering system holds back any token sequence that starts with [ and looks like it might be a citation. It accumulates tokens until the pattern either completes (matched as a citation and rendered as a badge) or fails to match (released as plain text). This prevents flicker without adding noticeable latency.
What users ask
The questions span a wide range:
- >"What were the top three objections across all discovery calls this quarter?"
- >"Which conversations mentioned timeline concerns for Q2?"
- >"Summarize the key decisions made across all sprint plannings this month"
- >"What patterns do you see in the coaching sessions about leadership development?"
- >"Compare the technical architecture discussions from January versus March"
The system handles both factual retrieval ("who said what") and synthesis ("what patterns exist"). Synthesis questions produce longer, more analytical answers. Factual questions produce shorter answers with more citations.
Persistent chat history
Folder chat maintains persistent history. Messages are stored per folder, organized by session. Users can start a new session to explore a different thread without losing previous conversations.
The chat context includes the 10 most recent messages, giving the LLM enough history for multi-turn conversations:
- >"Which prospects mentioned budget constraints?"
- >"Of those, which ones also expressed urgency?"
- >"Draft a follow-up strategy for the urgent ones"
Each turn builds on the previous context, narrowing and refining the search across the folder's conversations.
The indexing challenge
Vector indexing is not instant. A folder with 50 conversations, each with a detailed summary and multiple lenses, produces a significant amount of text to index.
We index lazily: the first time a user opens the chat for a folder, we check which conversations are not yet indexed and index them on the fly. Subsequent questions use the existing index. When a conversation is updated (new lens generated, summary regenerated), we bump a version number and re-index on the next query.
This version system ensures the index stays current without requiring background reindexing jobs. If the schema changes (we add new fields to the summary structure), we increment a global index version, and all conversations are re-indexed on next access.
What this enables
Folder intelligence changes how professionals interact with their conversation history.
A sales manager with a quarter of pipeline calls can identify patterns across the entire pipeline without reading individual notes. A product leader with a month of user interviews can synthesize themes without manually tagging and sorting. A coach with a year of client sessions can track developmental arcs across dozens of conversations.
The key insight: the value of recorded conversations scales non-linearly with the ability to search across them. Ten conversations, individually accessible, are ten separate documents. Ten conversations, collectively searchable, are a knowledge base.
Folder intelligence is the feature that makes the second case real.
