Skip to content

strategy.md

Latest commit

 

History

History
46 lines (35 loc) · 2.86 KB

strategy.md

File metadata and controls

46 lines (35 loc) · 2.86 KB
Raw

Codexa Strategy

Codexa's working beachhead hypothesis is the local-first researcher or archivist with a private 50k-200k-file corpus, a single workstation, and a need for cited answers rather than only file hits. Internal controls exercise that shape; external validation remains open in PMF-1.

Strategy Kernel

element statement
diagnosis Private corpora are too large for manual search, too sensitive for default cloud upload, and too heterogeneous for a single exact-match tool.
guiding policy Protect the local indexing -> retrieval -> grounded-answer loop first; keep every enrichment behind an opt-in adapter seam.
coherent actions Harden ingestion/retrieval quality gates, expose operator-visible evals, and sequence enrichments only when they improve the core jobs without adding default egress.

Durable-Advantage Hypothesis

The defensible bet is the combination of local-first privacy, swappable vendor seams, and cited grounding. Cloud RAG products can move faster on hosted model UX; Codexa's proposed advantage is that an operator can keep the corpus, index, providers, and eval evidence under local control. No external cohort yet establishes that this advantage is durable.

Now Next Later

horizon outcome evidence
Now Keep the index, search, and grounded-answer core mechanically complete while testing whether a new operator can finish it in one sitting. Deterministic demo/E2E controls pass; real-operator usability remains untested.
Next Test whether the production-path repo-doc signal transfers to private corpora. Per-PR CI indexes the bundled seven-document corpus and gates the live MiniLM + BM25 + RRF + MMR path; the deterministic lexical check is labeled as a harness smoke. Paired private-corpus evidence remains pending.
Later Promote optional enrichments only after measured job-outcome improvement without weakening the offline default. ZIM, vision, query enrichment, and online value evidence remain pending or rejected for wider defaults.

Sequenced Bets

  1. Make the core loop complete: index -> search -> grounded answer -> inspect citations.
  2. Make quality measurable: eval summaries, objective metric gaps, and regression gates.
  3. Add enrichment where it closes a measured gap: Wikipedia/ZIM, object detection, query expansion, and multi-turn UX stay subordinate to the core loop.

Validation Basis

Discovery notes and the evidence ledger live in ADR 0002. Only a complete row with an adopt decision, a predeclared threshold, and reproducible evidence may widen a default or move an enrichment beyond experimental use; pending and reject rows cannot be promoted. The full decision-record index is docs/adr/, and the no-default-egress posture is ADR 0004.