Buckets:
| Models & config | |
| 1. Which GGUF embedding model, and its dimension? If it isn't 768 the Qdrant collection gets recreated (it also needs recreating anyway for sparse vectors). GEMMA EMBEDDING 2 768 dim | |
| 2. Which OpenAI model(s)? One model throughout, or tiered — a cheap one for criterion classification/segmentation and a stronger one for entailment? Must be configurable. Azure openai only, set through config | |
| 3. Any cost ceiling per run? With ~87 trials fully verified this is the number that decides whether we batch criteria per call or go one-call-per-criterion. Decide later | |
| 4. OpenAI direct or Azure OpenAI? Key goes in the same backend/.env? Yes. | |
| 5. Confirm the venv is Python 3.13 (your system Python is 3.11). YES venv is 3.13 | |
| Pipeline semantics — I need exact rules, not my guesses | |
| 1. Tier boundaries: what exactly makes a trial high vs quick_check vs manual_review vs ineligible? Decision based on score ranges. | |
| 2. Trial verdict derivation: when is a trial eligible vs eligible_pending_unknown vs ineligible? Same as above. Decided by LLM | |
| 3. unknown vs unmet: the precise rule, especially for exclusion criteria where we have no evidence either way. Did not understand your question. | |
| 4. The cost table: do you want to supply the minute values, or should I propose a table in the TRD for you to approve? Table please | |
| 5. Is "87" a real config value? How many candidates come back from retrieval before the verified set — e.g. retrieve 300, verify 87? It will be dynamic based on retrieval and filtering. | |
| 6. Criteria per trial: any cap? Some CT.gov eligibility blocks parse into 60+ criteria, which blows up cost on a single trial. Suggest some numbers. | |
| Clarification loop | |
| 12. What does "Ask another round" actually do? New questions from where — gaps that round 1's answers newly exposed, or lower-priority gaps held back from round 1? Both, decided by LLM dynamically | |
| 13. Confirm: unresolved after round 3 → criterion marked unverifiable, trial tagged. YES | |
| 14. Are questions generated for gaps across all 87 trials, or only the top N? 87 trials' worth of gaps could be hundreds of questions; the mock shows a handful. Top N for now then. We'll decide later. | |
| Intake & profile | |
| 15. The free-text "focus" field ("e.g. immunotherapy only, no chemo") — what should it do? Query expansion, a hard filter, or an instruction to the ranker? Open to anything. | |
| 16. Editing the profile on the confirm screen — does an edit invalidate anything downstream, or just update the profile before matching starts? Update mismatches or wrong parsing | |
| 17. Sample patients: keep the 3 hardcoded in the frontend, or serve them from the backend? Backend. | |
| 18. How is the amber "low confidence" flag on an extracted fact derived? It will be culmination of a decision model and LLM (agent). New models have come to market like Jev, Laya etc. that are fast and give decisions. Look up the internet. | |
| API & frontend wiring | |
| 19. Browser → FastAPI directly with CORS, or proxied through the TanStack Start server? Direct. | |
| 20. Can a user cancel mid-run, and does the backend actually stop working? YES | |
| 21. Error contract — do you want a trace ID surfaced in the UI (the design doc's error state mentions one)? YES | |
| 22. Any auth or rate limiting at all, given local-only? No auth now | |
| Process | |
| 23. Test expectations: unit tests on predicates/parsers/cost model, integration against a fixture corpus, both, neither? CI? NO TEST, NO CI | |
| 24. Where should the TRD live — TECHNICAL_REQUIREMENTS.md in root? Any length or format preference? As long as required, but not too long. | |
| 25. Delivery order: thin end-to-end slice first (weak components, real wiring), or each stage built properly in sequence? | |
| 26. MILVUS_TOKEN/MILVUS_BASE_URL in .env — dead, delete them? YES. Sorry forgot to mention. Local embeddings storage only. FAISS. | |
| 27. This isn't a git repo. Want it initialized before we start? I'll do it later. | |
| Few things: I want to use Decision models locally for few parsing and scoring parts of the pipeline. Agents will take care of the rest more intelligent processing parts. |
Xet Storage Details
- Size:
- 4.09 kB
- Xet hash:
- e639210c14e60701e935deb0c47149ca7acd6a75ee39dda33eda2b25257bfa40
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.