Search conversations

Use the search button immediately beside the conversation-panel expand or collapse control to find text across your conversations. On mobile, use the equivalent search shortcut. You can also press Ctrl+K or Command+K while focus is outside an editable field. Opening search places keyboard focus in the query field so you can type immediately.

Choose the search scope#

Search starts with All conversations. Switch to This conversation to restrict results to the active graph. Results are grouped by conversation for scanning while retaining their overall relevance order.

Convography searches conversation and node titles, stored tags, user and assistant text, stored summaries, and structured decision metadata. It does not search profile information, file names, or text inside uploaded files.

Write a lexical query#

Search uses PostgreSQL full-text matching, not embeddings or semantic similarity. It matches words and their language stems; it does not infer unstated synonyms or meaning.

• Put a phrase in quotes to require adjacent words, such as "pricing model".
• Use OR between alternatives, such as postgres OR firestore.
• Prefix a word with - to exclude it, such as database -sqlite.
• A stored tag may appear as a matched field, but a tag is simply a short metadata label attached to a conversation or message by an existing workflow. Search does not add a tag editor.

The specific queries How did we get here?, Show the path here, and Show context for this node return the selected node's ancestor path. Select a graph node first; without one, search asks for an anchor. Other question-shaped queries are ordinary lexical searches.

Read and open results#

Results are ordered by relevance. Strong title and exact-tag matches rank above body matches; quoted phrases receive a boost, and recency only breaks otherwise close scores. Each result shows its conversation, node, safe text snippet, and the fields that matched.

Choose a result to switch conversations if needed, open its stable graph link, expand collapsed ancestors, select and focus the node, center it, and scroll its chat path to the source. Search does not place labels or badges over matching graph nodes. The chosen node uses the same node-and-edge relationship emphasis as any other clicked node; search does not add a separate primary-result or patterned ancestor treatment.

New or edited content is indexed in the background and normally appears within a few seconds. Active conversations that predate search enablement are also discovered and indexed automatically in bounded batches. While the index is building, an open search refreshes automatically as those batches finish. Pending deletions disappear immediately even if another index job is waiting.

Each time you open conversation search, the query field starts blank. Closing and reopening search does not restore the previous query.

Conversation search can cross conversations or be restricted to the active conversation. It does not search uploaded files and is not semantic search.