IntelliRAG. a pilot for answers worth trusting
INTELLIRAG / DESIGN PARTNER PILOT / SEPTEMBER 2026
Support answers.
With a trail you can trust.
Help technical teams find, check and reuse answers from their own documentation.
One source
→A cited answer
→A visible reason
Built by Charan Rathore · Working prototype · Seeking a measured pilot
THE COST HIDES BETWEEN THE QUESTION AND THE ANSWER
The docs exist.
The interruption still happens.
Find it again
Recurring questions pull engineers away from product work.
Described by the founders of YC-backed kapa.ai.
Check the version
A plausible answer can use the wrong configuration or an old release.
CircleCI’s support and schema-grounding use cases.
Know what applies
Buried updates and role-specific context slow distributed support teams.
Guru’s Perk customer account.
Public evidence motivates the pilot. These companies are not IntelliRAG customers; their outcomes are not our results.
START NARROW ENOUGH TO PROVE VALUE
One product.
One support queue.
The first team
A developer-facing SaaS or API business with changing docs and recurring setup questions.
Support lead owns the outcome. Agents and developer advocates use the tool.
The first question
“What does our source actually say about this behavior?”
Begin with public GitHub documentation. Use the team's real historical questions.
Target segment is a hypothesis. Interview five users before expanding to private, multilingual or company-wide search.
THE WORKING EXPERIMENT
Watch trust become inspectable.
Import
one issue
→Ask
within scope
→Trace
source + line
→Challenge
the unknown
→Reuse
the cache
The graph connects an answer to its evidence. A correction can guide retrieval; it never becomes a source fact.
Watch the real 57-second experiment ↗Public demo: keyword retrieval and cited extracts. Temporary imports; no language-model generation in this recording.
EVIDENCE BEFORE SALES CLAIMS
Small tests. Clear boundaries.
PASSEDTwo previously unseen READMEs
p-debounce and p-throttle: scoped citations, cited-only graph links, source excerpts, cached repeats and unsupported-question refusals.
Local tests plus a separate public issue walkthrough. This is fixture evidence, not customer accuracy.
OPEN WORKReliability beyond the happy path
GitHub tree requests returned 403; direct README imports passed. Private workspaces, automatic sync and durable production storage still need acceptance.
No claimed customer ROI, cost reduction or production-model accuracy.
A TWO-WEEK PILOT, AFTER SETUP
Bring 100 real questions.
01
Establish the baseline
Choose 20–50 documents. Time the current search and checking process. Hold back 30 questions.
02
Compare blindly
Tune on the training set. Two reviewers judge held-out answers and citations, without knowing the method.
03
Break it deliberately
Change or delete sources. Ask unsupported questions. Check cache invalidation and recovery.
Deliverable: a verified-answer scorecard, failure analysis and go/no-go decision. Public or approved non-confidential data first.
AGREE THE SCORECARD BEFORE THE PILOT
Measure useful answers.
| Outcome | Proposed target. not achieved |
|---|
| Time to a verified answer | 30% lower median; report p95 and quality together |
| Claims supported by citations | ≥95% supported; report answer coverage separately |
| Unknown or stale information | Zero invented high-risk answers; zero stale answers after refresh |
| Cost per verified answer | Measure actual provider usage and cache benefit |
| Later private-data access | Zero unauthorized evidence in cross-user / tenant tests |
Targets are negotiable pilot gates. Report sample sizes and reviewer disagreement. No generalized accuracy claim from a small test set.
BUILD WHAT REDUCES BUYER RISK
The next increments are practical.
Fresh and recoverable
Source revisions, last successful sync, rate-limit retries and incremental refresh.
Current imports refresh on request.
Persistent and permitted
Postgres, authenticated workspaces, source permissions and revocation tests.
Corpus filtering is not access control.
Correctable by an owner
A failed-question queue and review history. Learn whether corrections improve held-out answers.
Current feedback guides retrieval.
Kapa, Guru and Onyx already serve this market. Our hypothesis to test: faster evidence inspection and correction for a focused support workflow.
THE DESIGN PARTNER ASK
Bring the questions
your team keeps answering.
One owner. One workflow. A shared definition of a useful answer.
You bring
A representative question set, approved sources and a reviewer who knows the product.
We deliver
A scoped setup, transparent failure analysis and a measured recommendation, even if it is no-go.
Agree scope, acceptance criteria and fee before starting a paid pilot.
RESEARCH NOTES / PRIMARY SOURCES / 15 SEPTEMBER 2026
The evidence behind the direction.
Public research supports hypotheses; it does not rank pain across every startup or multinational. Vendor customer accounts are not independent studies. Full interview guide, test details and pilot protocol are in the repository.