root/moteva
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(research): hydrate web-search results with cleaned page content

Browse Source 
Fetch and clean each search result's page body (bounded by size,
concurrency, and a configurable MaxPageContentRunes), store it on the
new ResearchResult.Content field, and surface those excerpts in agent
memory so the model reasons over real page text rather than snippets
alone. Promotes golang.org/x/net to a direct dependency for HTML
parsing.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
liangxu
2026-07-08 10:56:31 +08:00
parent 222e5a0219
commit 836de5b597
8 changed files with 252 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
server/API.md
+2 -2
View File
@@ -171,7 +171,7 @@ Returns all project summaries, sorted by `updatedAt` descending.
}
```
`enableWebSearch` only opens the web-search capability. The agent model still decides per request whether search is actually needed; if it is not needed, no web-search step is recorded.
`enableWebSearch` only opens the web-search capability. The agent model still decides per request whether search is actually needed; if it is not needed, no web-search step is recorded. When search runs, result titles, URLs, snippets, and cleaned page-content excerpts are stored for the agent.
`GET /api/projects/:id`
@@ -210,7 +210,7 @@ Deletes the project record. PostgreSQL cascades canvas nodes, chat messages, his
}
```
The harness builds short memory from the current turn, recent messages, and visible canvas state, and long memory from durable project decisions, prior artifacts, and previous research records. Image titles and assistant language are generated from the request context instead of fixed model labels.
The harness builds short memory from the current turn, recent messages, and visible canvas state, and long memory from durable project decisions, prior artifacts, and previous research records with page-content excerpts. Image titles and assistant language are generated from the request context instead of fixed model labels.
## Generate