Add saved-tab question console mock

.github/workflows/idric-core.yml

@@ -29,6 +29,9 @@ jobs:
       - name: Exercise filesystem category and hot views
         run: sh bin/ci_browser_foundation.grease exercise-filesystem-views
 
+      - name: Exercise saved-tab text question mock
+        run: sh bin/ci_browser_foundation.grease exercise-tab-qa
+
       - name: Install system dependencies through Grease
         run: sh bin/ci_browser_foundation.grease install-dependencies
 

README.md

@@ -18,6 +18,8 @@ The browser core owns resource, tab, event, and task identity; sleeping and waki
 - `docs/filesystem-views.md` — category links, `_active`, and the requested `hot/` presentation set
 - `docs/storage-model.md` — identity levels and canonical, proposed, and derived state
 - `docs/developer-workbench.md` — fixture and memory-pressure harness
+- `docs/vector-index.md` — readable multi-model vector views and rebuildable Float32 query caches
+- `docs/tab-qa-mock.md` — console questions over saved reading pages with replaceable processing stages
 - `experiments/category-hyperplanes/README.md` — disposable embedding and explicit affine-separator probe
 - `experiments/embedding-models/README.md` — pinned tiny ONNX models and an end-to-end filesystem-index comparison
 

bin/ask_saved_pages.grease

@@ -0,0 +1,30 @@
+#!/bin/sh
+set -eu
+
+repository_root=$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd)
+export IB_TAB_QA_REPOSITORY_ROOT="$repository_root"
+. "$repository_root/lib/tab_qa.grease"
+
+case "${1:-}" in
+  --index)
+    test "$#" -eq 1 || exit 2
+    ib_tab_qa_build_index
+    ;;
+  --inspect)
+    test "$#" -eq 1 || exit 2
+    ib_tab_qa_inspect
+    ;;
+  --help)
+    printf '%s\n' \
+      "usage: $0 [--index|--inspect|QUESTION]" \
+      'With no argument, read one question from the console.'
+    ;;
+  '')
+    printf 'question> ' >&2
+    IFS= read -r question || exit 1
+    ib_tab_qa_answer "$question"
+    ;;
+  *)
+    ib_tab_qa_answer "$*"
+    ;;
+esac

bin/ci_browser_foundation.grease

@@ -74,6 +74,19 @@ exercise_filesystem_views() {
   sh tests/test_filesystem_views.grease
 }
 
+exercise_tab_qa() {
+  cd "$repository_root"
+  sh -n lib/tab_qa.grease
+  sh -n bin/ask_saved_pages.grease
+  sh -n bin/mock_tab_model.grease
+  sh -n bin/mock_tab_qa_reducer.grease
+  sh -n tests/test_mock_tab_model.grease
+  sh -n tests/test_tab_qa_pipeline.grease
+  make -C native/vector-index
+  sh tests/test_mock_tab_model.grease
+  sh tests/test_tab_qa_pipeline.grease
+}
+
 exercise_workbench() {
   cd "$repository_root"
   sh -n tests/test_real_world_url_fixture.grease
@@ -199,11 +212,12 @@ case "${1:-}" in
   exercise-workbench) exercise_workbench ;;
   exercise-file-store) exercise_file_store ;;
   exercise-filesystem-views) exercise_filesystem_views ;;
+  exercise-tab-qa) exercise_tab_qa ;;
   exercise-scientific-media) exercise_scientific_media ;;
   exercise-live-arxiv) exercise_live_arxiv ;;
   exercise-live-arxiv-prepaint) exercise_live_arxiv_prepaint ;;
   *)
-    printf 'usage: %s {install-dependencies|build-idric|verify-pdf-harvester|exercise-information|exercise-core|exercise-workbench|exercise-file-store|exercise-filesystem-views|exercise-scientific-media|exercise-live-arxiv|exercise-live-arxiv-prepaint}\n' "$0" >&2
+    printf 'usage: %s {install-dependencies|build-idric|verify-pdf-harvester|exercise-information|exercise-core|exercise-workbench|exercise-file-store|exercise-filesystem-views|exercise-tab-qa|exercise-scientific-media|exercise-live-arxiv|exercise-live-arxiv-prepaint}\n' "$0" >&2
     exit 2
     ;;
 esac

bin/mock_tab_model.grease

@@ -0,0 +1,116 @@
+#!/bin/sh
+set -eu
+
+model_name=mock-token-i8-v1
+dimensions=32
+
+inspect_model() {
+  printf '%s\n' \
+    "model=$model_name" \
+    'available=True' \
+    'model-kind=deterministic-token-test-double' \
+    'llm=False' \
+    'quantization=int8-counts' \
+    "dimensions=$dimensions" \
+    'tasks=embedding,extractive-answer'
+}
+
+embed_text() {
+  awk -v dimensions="$dimensions" '
+    function token_slot(token,    alphabet, character, hash, index_in_alphabet, position) {
+      alphabet = "abcdefghijklmnopqrstuvwxyz0123456789"
+      hash = 0
+      for (position = 1; position <= length(token); position++) {
+        character = substr(token, position, 1)
+        index_in_alphabet = index(alphabet, character)
+        hash = (hash * 37 + index_in_alphabet) % dimensions
+      }
+      return hash + 1
+    }
+    {
+      text = tolower($0)
+      gsub(/[^[:alnum:]]+/, " ", text)
+      token_count = split(text, tokens, / +/)
+      for (token_index = 1; token_index <= token_count; token_index++) {
+        token = tokens[token_index]
+        if (length(token) < 2)
+          continue
+        slot = token_slot(token)
+        if (values[slot] < 127)
+          values[slot]++
+      }
+    }
+    END {
+      for (slot = 1; slot <= dimensions; slot++)
+        printf "%s%d", slot == 1 ? "" : " ", values[slot] + 0
+      printf "\n"
+    }
+  '
+}
+
+answer_from_document() {
+  question=$1
+  document=$2
+  test -s "$document" || {
+    printf 'mock tab model: source document is empty: %s\n' "$document" >&2
+    return 1
+  }
+
+  awk -v question="$question" '
+    BEGIN {
+      normalized_question = tolower(question)
+      gsub(/[^[:alnum:]]+/, " ", normalized_question)
+      question_count = split(normalized_question, question_tokens, / +/)
+      for (question_index = 1; question_index <= question_count; question_index++)
+        if (length(question_tokens[question_index]) >= 2)
+          wanted[question_tokens[question_index]] = 1
+    }
+    NF {
+      if (first_line == "")
+        first_line = $0
+      normalized_line = tolower($0)
+      gsub(/[^[:alnum:]]+/, " ", normalized_line)
+      line_count = split(normalized_line, line_tokens, / +/)
+      score = 0
+      for (line_index = 1; line_index <= line_count; line_index++) {
+        token = line_tokens[line_index]
+        if (token in wanted && seen_on_line[token] != NR) {
+          score++
+          seen_on_line[token] = NR
+        }
+      }
+      if (best_line == "" || score > best_score) {
+        best_line = $0
+        best_score = score
+      }
+    }
+    END {
+      if (best_line != "")
+        print best_line
+      else if (first_line != "")
+        print first_line
+    }
+  ' "$document"
+}
+
+case "${1:-}" in
+  inspect)
+    test "$#" -eq 1 || exit 2
+    inspect_model
+    ;;
+  embed)
+    test "$#" -eq 1 || exit 2
+    embed_text
+    ;;
+  answer)
+    test "$#" -eq 3 || {
+      printf 'usage: %s answer QUESTION DOCUMENT\n' "$0" >&2
+      exit 2
+    }
+    answer_from_document "$2" "$3"
+    ;;
+  *)
+    printf 'usage: %s {inspect|embed|answer QUESTION DOCUMENT}\n' "$0" >&2
+    exit 2
+    ;;
+esac

bin/mock_tab_qa_reducer.grease

@@ -0,0 +1,68 @@
+#!/bin/sh
+set -eu
+
+reducer_name=mock-single-member-weighted-reducer-v1
+
+inspect_reducer() {
+  printf '%s\n' \
+    "reducer=$reducer_name" \
+    'available=True' \
+    'ensemble=single-member-test-double' \
+    'member-count=1' \
+    'xgboost=False'
+}
+
+reduce_candidate() {
+  retrieval_score=$1
+  source_document=$2
+  report=${IB_TAB_QA_REDUCER_REPORT:?IB_TAB_QA_REDUCER_REPORT must name the reducer report}
+  candidate=$(sed -n '1p')
+
+  test -n "$candidate" || {
+    printf 'mock reducer: model candidate is empty\n' >&2
+    return 1
+  }
+  test -s "$source_document" || {
+    printf 'mock reducer: source document is unavailable\n' >&2
+    return 1
+  }
+  awk -v value="$retrieval_score" 'BEGIN {
+    numeric = value ~ /^[-+]?([0-9]+([.][0-9]*)?|[.][0-9]+)([eE][-+]?[0-9]+)?$/
+    exit !numeric
+  }' || {
+    printf 'mock reducer: retrieval score is not numeric\n' >&2
+    return 1
+  }
+
+  printf '%s\n' \
+    "reducer=$reducer_name" \
+    'ensemble-stage-ran=True' \
+    'member-count=1' \
+    'checks-run=3' \
+    'candidate-nonempty=True' \
+    'source-readable=True' \
+    'score-numeric=True' \
+    "aggregate-score=$retrieval_score" > "$report"
+
+  # This one-member reducer deliberately preserves the model text. A later
+  # voting, bagging, rank-aggregation, or XGBoost adapter occupies this stage.
+  printf '%s\n' "$candidate"
+}
+
+case "${1:-}" in
+  inspect)
+    test "$#" -eq 1 || exit 2
+    inspect_reducer
+    ;;
+  reduce)
+    test "$#" -eq 3 || {
+      printf 'usage: %s reduce RETRIEVAL_SCORE SOURCE_DOCUMENT < candidate.txt\n' "$0" >&2
+      exit 2
+    }
+    reduce_candidate "$2" "$3"
+    ;;
+  *)
+    printf 'usage: %s {inspect|reduce RETRIEVAL_SCORE SOURCE_DOCUMENT}\n' "$0" >&2
+    exit 2
+    ;;
+esac

docs/tab-qa-mock.md

@@ -0,0 +1,84 @@
+# Saved-tab question-and-answer mock
+
+This is the first executable vertical slice of the ChatGPT-like text frontend.
+It is a console prompt, not the eventual phone UI, and it answers only from
+plain-text documents deliberately present below `~/reading`.
+
+The stages are explicit:
+
+1. `mock-token-i8-v1` tokenizes text into a 32-component, bounded 8-bit-count
+   test vector and selects one source line as an extractive answer.
+2. `ib-vector-index` stores one exact cosine index in the existing model-view
+   tree. Authoritative fixed-width text and its rebuildable Float32 cache are
+   two query paths over that index. Each query can record whether its dot
+   products used a GPU; the current C backend records `False` and `compute=cpu`.
+3. `mock-single-member-weighted-reducer-v1` is a one-member ensemble test
+   double. It runs three validations, records an aggregate score, and preserves
+   the candidate. This is the replaceable slot for voting, bagging, rank
+   aggregation, XGBoost, or another reducer; it does not claim to run XGBoost.
+4. `ib_tab_qa_null_render` is an identity post-processing step before the text
+   is returned.
+
+The bundled model is intentionally not described as an LLM. It is a tiny,
+queryable deterministic mock that exercises token, embedding, retrieval,
+model-output, reducer, and renderer boundaries without a model download.
+`IB_TAB_QA_MODEL_COMMAND` can name a later adapter implementing `inspect`,
+`embed`, and `answer`. `IB_TAB_QA_REDUCER_COMMAND` can replace the one-member
+reducer. The two pinned Hugging Face embedding manifests remain available for
+real model views; this mock writes `model-adapter.txt`, never a misleading
+`embedding-model.txt` manifest. A model name is bound to one exact adapter
+record, including the adapter command's SHA-256, so changed adapter code cannot
+silently reuse old vectors. An adapter with external model or vocabulary files
+must report their immutable hashes from `inspect` as part of the same record.
+
+## Filesystem boundaries
+
+Saved source text stays below `${IB_READING_DIR:-~/reading}`. The rebuildable
+index defaults to:
+
+```text
+${XDG_DATA_HOME:-~/.local/share}/ib/views/
+  organizing-the-information/vector-spaces/mock-token-i8-v1/reading/
+```
+
+Successful questions are appended as one directory per exchange below the
+separate requested folder:
+
+```text
+~/questions and answers about tabs that the user has visited/
+  <UTC-time>-<process>/
+    question.txt
+    model-candidate.txt
+    reduced-answer.txt
+    answer.txt
+    source.tsv
+    vector-query.tsv
+    vector-format.txt
+    reducer.tsv
+    indexing.tsv
+    model.tsv
+    reducer-model.tsv
+    reading-source.tsv
+    pipeline.tsv
+```
+
+The exchange preserves the raw model candidate, reduction evidence, selected
+source identity and content hash, upstream reading-source record, exact vector
+generation, adapter and reducer command hashes, vector execution report, and
+final response separately. It snapshots the selected document for processing
+and fails closed if those bytes differ from the indexed corpus record. The
+snapshot is temporary: canonical reading text is neither moved nor duplicated
+into the Q&A store. The command rejects a Q&A root equal to or nested with the
+reading root.
+
+## Run the mock
+
+```text
+make -C native/vector-index
+sh bin/ask_saved_pages.grease --index
+sh bin/ask_saved_pages.grease "What do feed-forward networks do to input space?"
+```
+
+Running the last command without a question displays `question>` and reads one
+line from the console. The index is reused until `--index` is run again; corpus
+change detection is deliberately left for the next slice.

docs/vector-index.md

@@ -36,6 +36,12 @@ document-id<TAB>0.8125
 
 `query` memory-maps the Float32 cache. `query-text` performs the same exact scan by parsing the readable file and works with the cache removed. `column` uses the fixed-width layout to print one coordinate across every ID. `compile-cache` atomically recreates the Float32 file from text. A later exact or approximate program can occupy the same process boundary without changing canonical browser state or callers that stream rows and queries.
 
+When `IB_VECTOR_QUERY_REPORT` names a file, either query command also writes a
+small execution report containing the compute device, storage path, metric,
+dimensions, and number of row dot products. The current portable C scan reports
+`dot-product-used-gpu=False`; a future accelerated backend must report its own
+execution rather than letting callers infer it.
+
 ## Files
 
 One model-specific view lives below the derived view root: