bastion-agent/docs/embed.md

bastion agent chunk and embed stages, and search

chunk splits the sanitized text into overlapping windows, embed turns each window into a vector with the model server EMBED_URL names (by default the Ollama port on the same machine), and search answers a query against those vectors. They are bastion_agent/embed.py, standard library only.

chunk

python3 -m bastion_agent chunk <manifest> [--out DIR] [--require-signed] reads the sanitized store (never the extract store) and writes chunks.jsonl: one row per window (chunk_id, doc (the original file's sha256, the id the extract and sanitized stores share), idx, start and end character offsets, chars, tokens, v), sorted by (doc, idx), plus chunk.summary.json.

• Windows are embed.chunk.size_tokens long and advance by size - overlap tokens; the last window may be shorter. With any overlap, every character of the text is inside at least one window (with overlap_tokens: 0 the whitespace between two windows may fall outside both).
• A "token" is a fixed regular-expression unit (a word or a punctuation mark), not a model's tokenizer. The standard library has none, and the same text must chunk the same way on every machine; the count is close enough to model tokens to size windows by.
• chunk_id = sha256(document sha256, NUL, window text). An unchanged document keeps its chunk ids, and so its vectors, from one sweep to the next. A changed document has a new sha256, so all of its windows get new ids and are embedded again. Identical windows inside one document (a repeated paragraph) share an id and are embedded once.
• Duplicate documents (the same file content at several paths) are chunked once.

embed

python3 -m bastion_agent embed <manifest> [--out DIR] [--require-signed] writes vectors.db (SQLite) and embed.summary.json.

The endpoint is a sensor. It comes from the environment, never the manifest: EMBED_URL (default http://127.0.0.1:11434, a local Ollama), EMBED_API (ollama or openai; unset means detect it with the probe), EMBED_BATCH (32), EMBED_TIMEOUT (300 s). The manifest pins the model name and dimension. Two wire formats: Ollama's POST /api/embed and the OpenAI-style POST /v1/embeddings (llama.cpp server, TEI).

The probe is the gate. Before the first real call the stage embeds a fixed probe string. If the served dimension is not the manifest's, the stage stops before opening the store, so a wrong or mislabelled model never writes a vector. An unreachable endpoint stops it the same way.

Storage (vectors.db):

Table	Role
chunks(chunk_id PK, doc, idx, start, end, text)	This sweep's windows, rebuilt on every run. Holds the sanitized window text, so a search can return it.
vectors(chunk_id, model, dim, vec BLOB float32, created), PK (chunk_id, model)	Kept across runs; rows are only ever added.

Switching models adds rows for the new model; the old model's rows stay. Vectors whose chunk no longer exists in this sweep also stay, and are counted as stale_vectors in the summary: nothing is thrown away, so going back to an earlier model or document costs nothing. A re-run embeds only chunk ids that have no vector for the manifest's model (the rest are reused).

search

python3 -m bastion_agent search <manifest> "<query>" [-k N] [--out DIR] embeds the query through the same endpoint, scores cosine similarity against every vector of the manifest's model, and prints the top k (default 5) sanitized windows with their score, document id and index. It is a linear scan in pure Python, which is fine to tens of thousands of chunks; an index (HNSW or sqlite-vec) is the next step for a folder that outgrows it. The query endpoint (serve, serve.md) uses the same search.

Tests

python3 -m unittest bastion_agent.test_embed runs against a loopback stub that speaks both wire formats with a deterministic bag-of-words embedding: window coverage, overlap and step; empty and short inputs; content-addressed ids; the float32 packing and cosine edge cases; the chunk stage (duplicate documents, sort order, determinism, the sanitize-first guard); the probe refusing a wrong dimension before the store exists; an unreachable endpoint failing closed; batching; reuse on a re-run; a second model adding rows while the first model's rows stay; stale vectors kept and counted after a document changes; OpenAI-style detection and a forced wrong API refused; and search ranking the right document, with an empty result for a model that has no vectors.