Datasets:
Upload ctxstream/README.md with huggingface_hub
Browse files- ctxstream/README.md +91 -0
ctxstream/README.md
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ctxstream
|
| 2 |
+
|
| 3 |
+
Stream an oversized context past a small model, then construct the answer in code.
|
| 4 |
+
|
| 5 |
+
Zero third-party dependencies. C++17.
|
| 6 |
+
|
| 7 |
+
## Why
|
| 8 |
+
|
| 9 |
+
Measured on a 4B model over a 261,226-token corpus on a 6GB card: the model swept
|
| 10 |
+
every segment correctly, was asked which label was least common, and answered
|
| 11 |
+
|
| 12 |
+
Status: beta, Status: delta, Status: gamma, Status: alpha
|
| 13 |
+
|
| 14 |
+
It listed the candidates instead of selecting one. An earlier run with 65 segments
|
| 15 |
+
combined the partial counts wrongly. Same mistake both times, and it was the
|
| 16 |
+
harness's, not the model's: a language model was asked to plan a traversal *and*
|
| 17 |
+
to do aggregation arithmetic.
|
| 18 |
+
|
| 19 |
+
ctxstream takes both jobs away from it.
|
| 20 |
+
|
| 21 |
+
| video streaming | ctxstream |
|
| 22 |
+
|---|---|
|
| 23 |
+
| manifest / playlist | segment plan, computed in code, before any model call |
|
| 24 |
+
| buffer | N segments in flight, latency hidden behind compute |
|
| 25 |
+
| decoder | model sees ONE segment, emits `key<TAB>number`, never prose |
|
| 26 |
+
| playback | reduce phase — a strategy you choose, in code |
|
| 27 |
+
|
| 28 |
+
The model's only job is extraction from a window it comfortably fits.
|
| 29 |
+
|
| 30 |
+
## Codebase mode
|
| 31 |
+
|
| 32 |
+
A directory input is not treated as a character stream. Cutting code every N chars
|
| 33 |
+
splits functions and separates a call from its definition. Instead:
|
| 34 |
+
|
| 35 |
+
1. **scan** — walk the repo, skip excluded dirs and binaries, record what was
|
| 36 |
+
skipped and why
|
| 37 |
+
2. **graph** — symbols (function/class/struct) per file, plus file→file edges from
|
| 38 |
+
`#include` / `import` / `require`
|
| 39 |
+
3. **order** — dependency-first, so a file follows what it includes
|
| 40 |
+
4. **segment** — on file boundaries, never mid-symbol; oversized files fall back to
|
| 41 |
+
line-boundary splitting
|
| 42 |
+
|
| 43 |
+
```bash
|
| 44 |
+
ctxstream --input ./my-repo --graph-only # inspect the graph
|
| 45 |
+
ctxstream --input ./my-repo --extract "..." # graph, then stream
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
## Build
|
| 49 |
+
|
| 50 |
+
```bash
|
| 51 |
+
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
|
| 52 |
+
cmake --build build -j
|
| 53 |
+
./build/test_ctxstream # 61 checks, no model or network needed
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
## Use
|
| 57 |
+
|
| 58 |
+
```bash
|
| 59 |
+
ctxstream --input corpus.txt \
|
| 60 |
+
--extract "Count how many lines carry each category value in this fragment." \
|
| 61 |
+
--answer least --label Category \
|
| 62 |
+
--model gemma4:e4b --num-ctx 32768 --segment-chars 60000
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
`--answer` is `least | most | total | list`. `--plan-only` prints the segment plan
|
| 66 |
+
and exits without calling a model.
|
| 67 |
+
|
| 68 |
+
## Guarantees it actually enforces
|
| 69 |
+
|
| 70 |
+
- **Full coverage.** Every byte lands in some segment; tests assert it. A planner
|
| 71 |
+
that drops the tail yields a confident answer over partial data.
|
| 72 |
+
- **No mid-record cuts.** Segments snap back to a record boundary.
|
| 73 |
+
- **Truncation is fatal, not silent.** A server that clips the prompt and answers
|
| 74 |
+
from the remainder is caught by comparing reported prompt tokens against what
|
| 75 |
+
was sent. Measured case: 50,000 tokens sent, 16,387 processed, no error raised.
|
| 76 |
+
- **A partial sweep refuses a headline answer.** If any segment failed, every key
|
| 77 |
+
is undercounted, so `--answer least/most/total` exits non-zero rather than
|
| 78 |
+
printing a number.
|
| 79 |
+
- **Keys with spaces survive.** `Spatial Relationship` must not become `Spatial`.
|
| 80 |
+
|
| 81 |
+
## Layout
|
| 82 |
+
|
| 83 |
+
```
|
| 84 |
+
src/ctxstream.hpp types + API
|
| 85 |
+
src/manifest.cpp segment planning (text)
|
| 86 |
+
src/codegraph.cpp repo scan, symbols, edges, graph-aware segmentation
|
| 87 |
+
src/backend.cpp HTTP/1.1 + minimal JSON, no libcurl
|
| 88 |
+
src/pipeline.cpp concurrent streaming, retries, progress
|
| 89 |
+
src/reduce.cpp record parsing, tally, answer construction
|
| 90 |
+
src/main.cpp CLI
|
| 91 |
+
```
|