Automated Research

[jaunty hearth]

retrieval/ranking logic is not properly prioritizing recency and relevance when selecting evidence sources.

[pliant cedar]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[frail quail]

Yeah, that sounds like a ranking bug / tuning issue rather than something users can fix from config alone.

Could you share:

• openclaw --version
• the exact research/query prompt
• 2–3 examples of sources it picked
• the newer/more relevant sources you expected it to prefer
• whether this is web search, memory/session retrieval, or document evidence retrieval

If it’s reproducible, please also file it here so it doesn’t get lost: https://github.com/openclaw/openclaw/issues

The most useful repro is “query → selected evidence → expected evidence,” especially with timestamps/URLs. That gives maintainers something concrete to tune or test against.

[jaunty hearth]

Current issue summary:

The agent is now successfully:

researching companies
extracting relevant operational information
identifying legitimate selling signals
writing durable outputs correctly

However, the evidence-selection logic still needs refinement.

Primary problem:

the agent surfaced a 2015 article as a primary supporting source
while semantically relevant, the information was too old to be operationally useful

What needs improvement:

stronger recency/freshness filtering
recency-weighted source ranking
rejection or downgrade of stale evidence
distinction between historical context vs active/current selling signals
require newer corroborating evidence before confirming a signal

Desired behavior:

prioritize sources from the last 12–24 months
prioritize recent expansions, capex, tooling, automation, hiring, EV programs, facility investments, etc.
if only old evidence exists, mark the signal weak/stale instead of active
do not allow outdated articles to become primary evidence for current opportunities

This appears to be a retrieval/evidence-ranking issue, not a research or automation issue.

[frail quail]

Yep — this is the right diagnosis. The agent is finding semantically relevant evidence, but it’s not treating evidence age as a first-class quality signal.

A good fix should probably make the evidence scorer do something like:

• hard-prefer evidence from the last 12–24 months for “active opportunity” claims
• allow old sources only as historical/company background
• downgrade anything older than ~24 months unless corroborated by newer evidence
• require at least one recent source before marking a signal as active/current
• label old-only signals as stale, weak, or historical, not as confirmed selling signals

For the agent instructions/output schema, I’d add a guardrail like:

For operational/selling signals, evidence older than 24 months must not be used as primary evidence unless corroborated by a newer source. If no recent corroboration exists, mark the signal as stale/historical and lower confidence.

And for the durable output, I’d include fields like: