EmbodiedEdge Labs · Evidence
DAS Field Commissioning Copilot · Distributed antenna systems Working on local hardware · cloud deploy next

When the manual doesn’t say, the copilot says so.

Closing the gap between the spec sheet and the site.

A distributed antenna system is the last physical layer of a cellular network, where the carrier's signal has to work inside a real building. Technicians installing and commissioning one ask the copilot what to do next. It answers from OEM manuals and general DAS field guides, keeps every value and safety warning as the source states it, and says plainly when no source covers the question instead of guessing.

ModelGemma 4 (e4b)Ollama on an RTX 5090 workstation
LibrarySeven manufacturersplus general DAS field guides
RetrievalLayered, scope-awareEmbeddings plus keyword index
Checks548 tests · 66/66unit tests, cross-OEM benchmark
The app

Two ways in

Front page of the DAS Field Commissioning Copilot app, offering User Mode and Training Mode.
Front page of the running app, captured October 3, 2026.
User Mode

The technician view. Ask a commissioning question and get field steps with a badge showing how well the sources back the answer, without the review tools in the way.

Training Mode

For the people who keep answers trustworthy. Review the references behind each answer, grade answers, and accept web-assisted answers into review. Opens with a power-user key.

User Mode chat screen with a question about a VSWR alarm on a DAS remote typed into the prompt box.
User Mode, ready for a question. The prompt box takes plain field language; every answer carries the reminder to verify before field execution.
Signal path

Where the questions come from

Each stage of a DAS has its own measurements and its own failure modes. These are questions the running system was asked on October 3, 2026, placed at the stage they belong to.

SIGNAL SOURCE BTS · BDA · off-air POI / HEADEND attenuation, dB FIBER TRANSPORT optical budget, dB REMOTE UNIT output, dBm · alarms COAX + ANTENNAS RL, dB · PIM, dBc How is TDTPOI attenuation set during commissioning? Fiber link to remote 3: low optical power. What to check? A DAS remote shows a VSWR alarm. Troubleshoot? PIM failed at -140 dBc on the sector 2 run. What to check? What is the end-to-end process to commission a DAS? DOWNLINK →
Schematic. Copper lines carry RF, blue lines carry light.Questions are verbatim from the live test session
Method, per question

How an answer is built

01 · readUnderstand the question

Finds the intent, the vendor and platform, and any measurement with its unit, such as 14 dB return loss.

02 · planDecide which sources count

Sets how many slots go to the platform's OEM manuals, general DAS guides, and project criteria.

03 · retrieveRank the evidence

Scores by relevance and authority, filters out other vendors, and pushes cover pages and legal notices down.

04 · writeAnswer in field language

Rewrites manual text into short action steps, keeping values, part numbers, menu paths and warnings exact.

05 · checkVerify before showing

Flags any number the sources do not contain and labels the answer fully grounded, partial, or not covered.

A missing source is an answer too

When no manual or guide covers a question, the copilot says so in one plain sentence and points to the document that would settle it. A confident guess about a torque value or an alarm meaning is worse than no answer on a live site.

Live answers · October 3, 2026 · abridged

Four ways it responds

Each card is a real question to the running system and the start of its answer. The badges are the ones the technician sees. In use, the tech names the manufacturer and platform, and the copilot answers only from that system’s documentation. Names are left out here.

Grounded in the OEM manual

troubleshooting
A DAS remote shows a VSWR alarm. How do I troubleshoot it?
Check the RF path components
  • Inspect the coax path for visible damage or stress.
  • Verify every connector along the path is mated and clean.
  • Check that the antenna connections are secure.
Check the system status
  • Open Command & Control, select the BIU, then review History > Event Log.
Fully groundedOEM 3General DAS 2

Why it reads this way: fault questions about a named platform keep two evidence slots for general fault isolation, so the physical checks come first and the vendor menus second.

General checks, clearly labeled

no platform procedures on file
The optical link to remote 3 shows low optical power. What do I check?
These are general DAS checks, not this platform’s procedures. Confirm alarm meanings, thresholds, and the approved procedure in the manufacturer’s documentation before acting.
  1. Inspect the connection, clean it, then inspect it again.
  2. Verify the wavelength, fiber, polish, and polarity.
  3. Measure both endpoint powers.
  4. Run OLTS or OTDR as appropriate.
  5. Compare the results with the engineered budget.
Partially groundedGeneral DAS 5

Why it reads this way: a tech asking what to check gets the method. Asking what the alarm means on this platform is still refused, because that answer has to come from the manufacturer.

No pass/fail without criteria

measurement
Is 14 dB return loss a pass?
I can explain what 14 dB means, but I cannot call it pass or fail without the applicable project, carrier, OEM, regulatory, standard, or AHJ acceptance criteria. Next checks
  • Find the approved acceptance criteria for the site, band, and measurement location before recording a result.
Partially groundedGeneral DAS 5

Why it reads this way: the same number can pass on one project and fail on another. The verdict belongs to the acceptance document, not the model.

Declines rather than guess

specification
What torque should I use on a 4.3-10 connector?
No verified guidance covers this, so no answer is given rather than a guess. Check the OEM documentation or the project acceptance criteria for this site.
Insufficient evidenceGeneral DAS 5

Why it reads this way: none of the sources states a torque value. A plausible number from the model's memory would read as fact on site.

Source precedence

Which document wins

When sources disagree, the copilot follows the order a commissioning lead would. General guides explain; they never overrule a project document or a manufacturer.

  1. 1Approved project or site procedure and acceptance criteriasite truth
  2. 2OEM manual for the exact platform, model and revisionproduct truth
  3. 3Carrier or customer engineering documentsnetwork owner
  4. 4Applicable regulations and standardscode
  5. 5Verified general DAS guidesmethod
  6. 6The model's own background knowledgelast resort
No cross-vendor answers

A question about one manufacturer’s remote never draws on another manufacturer’s manual. The benchmark checks this on every case and the contamination rate is zero.

Older revisions step aside

When a newer edition of a manual is in the library, the superseded one is marked deprecated and left out of answers.

Conflicts are shown, not settled quietly

If two sources of the same rank disagree on a value, the answer says so and asks for field verification.

Library

What it draws on

Installation, commissioning and operations guidance across seven manufacturers, plus general DAS guides covering fundamentals, design, fiber, coax, testing, troubleshooting and public safety.

Measured

How the answers score

0.906

Faithfulness: the share of each answer's claims an independent judge model traces to the retrieved evidence.

0

Unflagged numbers missing from the sources. Every unsupported value is marked for verification.

66/66

Cross-OEM retrieval benchmark cases passed, with no other-vendor evidence accepted.

9.5%

False refusals on questions whose source manual is in the library (2 of 21).

0.833

Decline recall: questions the library does not cover, declined instead of answered from model memory.

1.4 s

Typical response time on the workstation; 11.0 s at the 95th percentile.

Generation evaluation on the workstation: Gemma 4 e4b answering, Qwen3 30B judging, 26 judged answers. Questions whose expected manual is not in the library were scored separately, since declining is the correct answer for them. Repeat judging of the same answers moved faithfulness by about 0.04, so smaller differences are not claimed.

The number guard

Every number is sourced or flagged.

The first baseline showed the risk plainly: before the guard, a large share of the numbers with units in answers appeared nowhere in the retrieved evidence or the question. A prompt change alone did not fix it. A deterministic check after generation did: any number with a unit that the evidence does not contain is marked [UNVERIFIED - verify before field use], so the technician sees the warning instead of a clean-looking value.

Small model, Qwen3 1.7B

41.5%→0%

Unsourced numbers shown as fact: 17 of 41 before; 0 of 28 after, with 14 flagged for the technician.

Larger model, Qwen3 30B

37.0%→0%

Unsourced numbers shown as fact: 20 of 54 before; 0 of 41 after, with 3 flagged for the technician.

Same 65 cases, same judge, lexical corpus backend, before and after the prompt fix and number guard. The cross-OEM retrieval benchmark was 66/66 before and after.

Model choice

The default model was chosen by measurement.

Four configurations scored on the same cases with the same judge. The smallest model kept zero unsourced numbers only because the guard flagged 17 of them; Gemma 4 produced none that needed flagging and the highest faithfulness, so it became the default.

A negative result stays in the record: enabling the model’s thinking mode under the default token budget dropped faithfulness to 0.099, because thinking consumed the answer budget. The setting exists and stays off by default.

Field testing

What live testing found and what was fixed.

Running the copilot from a browser the way a technician would surfaced problems the offline scores missed. Each was fixed and the full evaluation re-run; faithfulness rose from 0.874 to 0.906 with no new unsourced numbers.

Found in the field UIFix
“How do I commission a DAS?” got a vendor clarification instead of stepsBroad questions get vendor-neutral steps first, then the platform-specific options
Quoted manual text came back lowercased and broken apartPassages are stored and shown exactly as the manual prints them; index rebuilt
Declines looked like a half-filled answer with a “grounded” badgeOne plain message that the sources don’t cover the question, with no badge
Answers repeated source wording; cover pages and regulatory boilerplate took top resultsAnswers rephrased in plain field language, keeping values, units, part numbers, alarm codes and menu paths exact; boilerplate ranked down
Citations showed the first lines of a page, not the relevant passageCitations show the sentences that match the question; the copilot never points to a figure the technician cannot see
Status, October 2026

What is proven, and what is next

Working now

  • Runs end to end on the workstation: API, chat interface, and a local model, tested live from a second machine on the network.
  • Step-by-step workflows for commissioning, with each step's measurement validated and every event recorded for handoff.
  • 548 automated tests and lint run on every push, on Python 3.11 and 3.12.
  • Model choice is a setting: local Ollama by default, with OpenAI-compatible and Amazon Bedrock clients ready.

Next

  • Cloud deployment on AWS App Runner with Amazon Bedrock, from the written runbook.
  • Re-score generation on the current build.
  • Field trial with working technicians on a live commissioning job.

Advisory by design. The copilot guides, validates, explains and records. It never logs into equipment, pushes configuration, or writes to operations systems. People execute the approved procedure.

Evidence

Where every number comes from.

  • Generation evaluation score log: every run, model, judge and configuration
  • Working local system handoff, 2026-10-03: the measured build and the field-testing fixes
  • Evaluation cases: 28 generation, 37 grounding and 66 cross-OEM retrieval cases

The source repository is private. Evaluation files and the score log are available on request.

An EmbodiedEdge Labs project. Measured on real hardware, published as found.