Skip to content

feedback-loop.md

Latest commit

 

History

History
107 lines (76 loc) · 3.83 KB

feedback-loop.md

File metadata and controls

107 lines (76 loc) · 3.83 KB

Feedback and Profile Learning

PaperFlow treats feedback as a shared learning signal. Whether the user replies through the CLI, local GUI, or Feishu/Lark bot, the signal is stored in the same local database and updates the same user profile for that user_id.

Feedback Entry Points

Surface User action Profile effect
CLI paperflow feedback --user-id ... --push-id ... --reply "1 3" Updates selected/skipped paper signals and drift state
GUI Select papers, mark "not interested", then submit Updates selected/skipped paper signals and drift state
Feishu/Lark bot Reply with numbers like 1 3, ranges like 1-5, or none Routes to the same feedback agent and updates the same profile
Reading report Run paperflow read, read an arXiv ID, or read a local PDF Adds a conservative reading-side interest signal
Reading-report quality feedback Send positive/negative report feedback in Feishu chat Updates report-style preferences in the profile

The important rule is that the user_id is the identity boundary. Feedback only updates the profile for the matching user or role.

What Gets Updated

Feedback can update:

  • topic_weights and other profile interest weights
  • selected/skipped behavior history
  • short-term and long-term drift state
  • reading signal state for papers that were deeply read
  • local wiki nodes/edges when PAPERFLOW_WIKI_INGEST=true
  • report-style preferences when the feedback is about reading-report quality

CLI Flow

paperflow daily --user-id user_alice
paperflow feedback \
  --user-id user_alice \
  --push-id push_20260601_090000 \
  --reply "1 3"

This records papers 1 and 3 as selected. On the first feedback for a push, the remaining unselected papers are treated as weak skipped signals. Later amendments only add newly selected papers, so a second reply does not erase the first selection.

Use:

paperflow feedback --user-id user_alice --push-id push_20260601_090000 --reply "none"

to explicitly say none of the pushed papers were useful.

GUI Flow

In paperflow gui, the daily-push panel has two explicit actions:

  • select papers to read
  • mark papers as not interested

Submitting either action records behavior logs, updates the profile, and refreshes local wiki metadata. If "提交并精读" is used, selected papers also generate reading reports after the feedback update.

Feishu/Lark Flow

In the webhook bot, chat messages such as 1 3, 1-5, none, 全部, and 没有 are parsed as feedback. The bot resolves the current role/user from the chat context and then calls the same feedback agent used by the CLI.

That means a Feishu reply and this CLI command are equivalent learning signals when they target the same user_id and push:

paperflow feedback --user-id user_alice --push-id <latest_push_id> --reply "1 3"

Feishu delivery requires webhook/ngrok only for receiving chat events. The profile update itself is local and stored in data/paperflow.db.

Reading Signals

Reading reports are also profile signals, but they are handled more conservatively than explicit selection feedback. When a user runs paperflow read, reads an arXiv paper in the GUI, or uploads/reads a local PDF, PaperFlow can reinforce the detected paper topics through the reading-signal state.

This helps the system learn from "I actually read this" without treating every deep read as a hard preference shift.

Local Wiki

When PAPERFLOW_WIKI_INGEST=true, feedback and drift updates are mirrored into the local wiki:

  • paper nodes record selected/skipped/read behavior
  • trajectory nodes record push and drift transitions
  • topic nodes connect repeated themes

Run:

paperflow wiki backfill --user-id user_alice
paperflow wiki stats --user-id user_alice

to import older runtime history and inspect the stored signal counts.