|
| 1 | +# Retrieval |
| 2 | + |
| 3 | +The retrieval capability covers the two building blocks of a search or |
| 4 | +RAG pipeline: turning text into vectors (**embedding**) and reordering a |
| 5 | +candidate list by relevance to a query (**reranking**). Like the LLM |
| 6 | +capability, both are async IO you call from inside a node body, behind a |
| 7 | +small provider protocol so the graph does not care which vendor is on the |
| 8 | +other end. |
| 9 | + |
| 10 | +Everything lives in `openarmature.retrieval`: |
| 11 | + |
| 12 | +```python |
| 13 | +from openarmature.retrieval import ( |
| 14 | + OpenAIEmbeddingProvider, |
| 15 | + CohereRerankProvider, |
| 16 | + EmbeddingRuntimeConfig, |
| 17 | + RerankRuntimeConfig, |
| 18 | +) |
| 19 | +``` |
| 20 | + |
| 21 | +## Embedding text |
| 22 | + |
| 23 | +An `EmbeddingProvider` turns a list of strings into a list of vectors, |
| 24 | +one vector per input, in input order: |
| 25 | + |
| 26 | +```python |
| 27 | +provider = OpenAIEmbeddingProvider(model="text-embedding-3-small", api_key="sk-...") |
| 28 | + |
| 29 | +response = await provider.embed(["the lunar south pole", "the Sea of Tranquility"]) |
| 30 | + |
| 31 | +response.vectors # list[list[float]], one per input, same order |
| 32 | +response.dimensions # the vector length (equal across all vectors) |
| 33 | +response.model # the model the provider actually served |
| 34 | +response.usage # an EmbeddingUsage record, or None (see below) |
| 35 | +``` |
| 36 | + |
| 37 | +The input-order guarantee is the contract you build on: `vectors[i]` is |
| 38 | +always the embedding of `input[i]`, no matter how the provider paginated |
| 39 | +the request under the hood. `len(response.vectors) == len(input)` always |
| 40 | +holds, and every vector has the same dimensionality. |
| 41 | + |
| 42 | +`embed` takes an arbitrary-length list. You do not have to pre-chunk to |
| 43 | +fit a provider's per-request cap; the mapping does that for you (see |
| 44 | +[Long input lists](#long-input-lists-are-chunked-for-you)). |
| 45 | + |
| 46 | +## Reranking candidates |
| 47 | + |
| 48 | +A `RerankProvider` scores a set of candidate documents against a query |
| 49 | +and returns them sorted best-first. This is the precision step after a |
| 50 | +cheap-and-broad first pass (vector similarity, keyword search, whatever): |
| 51 | + |
| 52 | +```python |
| 53 | +provider = CohereRerankProvider(model="rerank-v3.5", api_key="...") |
| 54 | + |
| 55 | +response = await provider.rerank( |
| 56 | + "where is there water ice on the Moon?", |
| 57 | + ["Regolith is abrasive dust.", "Ice sits in shadowed polar craters.", "..."], |
| 58 | + top_k=3, |
| 59 | +) |
| 60 | + |
| 61 | +for result in response.results: # sorted by relevance_score, best first |
| 62 | + result.index # position in the input `documents` list |
| 63 | + result.relevance_score # higher is more relevant |
| 64 | + result.document # the echoed text, or None (see below) |
| 65 | +``` |
| 66 | + |
| 67 | +Results come back ranked. The mapping applies the sort even if a provider |
| 68 | +returns them unsorted, so you can always trust `response.results[0]` to be |
| 69 | +the best hit. |
| 70 | + |
| 71 | +`result.index` is the load-bearing field: it points back into the |
| 72 | +`documents` list you passed. Not every provider echoes the document text |
| 73 | +(Cohere, for one, returns scores only), so `result.document` may be |
| 74 | +`None`. Map back to your own candidate list by index rather than relying |
| 75 | +on the echo: |
| 76 | + |
| 77 | +```python |
| 78 | +ranked = [candidates[result.index] for result in response.results] |
| 79 | +``` |
| 80 | + |
| 81 | +`top_k` trims the returned results to the best K. Pass it when you only |
| 82 | +want the top of the list; omit it to score every candidate. |
| 83 | + |
| 84 | +## Query vs document: `input_type` |
| 85 | + |
| 86 | +Some embedding models are **asymmetric**: they embed a search query and |
| 87 | +the documents being searched with slightly different representations, and |
| 88 | +mixing them up hurts recall. `EmbeddingRuntimeConfig.input_type` is the |
| 89 | +portable knob for this: |
| 90 | + |
| 91 | +```python |
| 92 | +# When embedding the corpus you are searching over: |
| 93 | +await provider.embed(passages, config=EmbeddingRuntimeConfig(input_type="document")) |
| 94 | + |
| 95 | +# When embedding the user's query at search time: |
| 96 | +await provider.embed([query], config=EmbeddingRuntimeConfig(input_type="query")) |
| 97 | +``` |
| 98 | + |
| 99 | +Each provider realizes it the way its wire expects: on an asymmetric |
| 100 | +model it selects the query or document representation; on a symmetric |
| 101 | +model (OpenAI's) it is a no-op and the text is embedded verbatim. Setting |
| 102 | +it costs nothing on the symmetric providers and keeps the same pipeline |
| 103 | +correct if you later switch to an asymmetric one, so prefer setting it. |
| 104 | + |
| 105 | +## Long input lists are chunked for you |
| 106 | + |
| 107 | +Every hosted embedding API caps how many inputs one request may carry. |
| 108 | +The mapping honors the contract that `embed` takes any-length input by |
| 109 | +splitting a large list into consecutive slices under the cap, issuing one |
| 110 | +request per slice, and stitching the vectors back together in input |
| 111 | +order. This is invisible: you call `embed` with 10,000 strings and get |
| 112 | +10,000 vectors, regardless of the provider's per-request limit. |
| 113 | + |
| 114 | +Usage is combined across the slices: the token total is reported only |
| 115 | +when every slice reported one, otherwise `usage` is `None` for the whole |
| 116 | +call (an honest "unknown" rather than a partial count). |
| 117 | + |
| 118 | +## Usage is a record or null |
| 119 | + |
| 120 | +`response.usage` is an `EmbeddingUsage` (or `RerankUsage`) record when the |
| 121 | +provider reported token accounting, and `None` when it did not. The |
| 122 | +mapping never fabricates a usage record, a zero, or a client-side |
| 123 | +estimate: a `None` means "the provider told us nothing," which is |
| 124 | +different from a real, reported zero. Guard before reading it: |
| 125 | + |
| 126 | +```python |
| 127 | +if response.usage is not None: |
| 128 | + print(response.usage.input_tokens) |
| 129 | +``` |
| 130 | + |
| 131 | +This matters because not every provider bills the same way. A local |
| 132 | +Text Embeddings Inference server reports no usage at all, so `usage` is |
| 133 | +`None`. Cohere's reranker reports `search_units` but no token count, so |
| 134 | +its `RerankUsage` carries `search_units` with `input_tokens=None`. |
| 135 | + |
| 136 | +## The bundled providers |
| 137 | + |
| 138 | +Four vendors ship as reference providers: OpenAI-compatible (embedding |
| 139 | +only), Cohere, Jina, and TEI. All four embed; Cohere, Jina, and TEI also |
| 140 | +rerank. The [Retrieval Providers](../retrieval-providers/index.md) section |
| 141 | +has the full table, the protocol contract, per-provider notes, and a guide |
| 142 | +to [self-hosting TEI](../retrieval-providers/tei.md) for embeddings and |
| 143 | +reranking on your own hardware. Writing your own is the same exercise as a |
| 144 | +custom LLM provider: implement the `EmbeddingProvider` or `RerankProvider` |
| 145 | +protocol and the graph treats it like any other. |
| 146 | + |
| 147 | +## Observability |
| 148 | + |
| 149 | +When you call `embed` or `rerank` from inside a node body, the provider |
| 150 | +dispatches a typed `EmbeddingEvent` or `RerankEvent` (and a failed |
| 151 | +variant on error) to any attached observer, exactly like LLM completions. |
| 152 | +The bundled OTel observer renders an `openarmature.embedding.complete` or |
| 153 | +`openarmature.rerank.complete` span; the Langfuse observer renders a |
| 154 | +dedicated Embedding or Retriever observation. Token usage lands on the |
| 155 | +span or observation only when the provider reported it, matching the |
| 156 | +nullable-usage contract above. |
| 157 | + |
| 158 | +Provider calls made outside a graph (for example an offline index build) |
| 159 | +run fine but dispatch no events, since there is no observer context to |
| 160 | +receive them. |
| 161 | + |
| 162 | +## Putting it together |
| 163 | + |
| 164 | +The [`retrieval-rag`](https://github.com/LunarCommand/openarmature-python/tree/main/examples/retrieval-rag) |
| 165 | +example wires the whole pattern into a graph: it batch-embeds a corpus |
| 166 | +into an index, then per query embeds the question, retrieves by cosine |
| 167 | +similarity, reranks the shortlist, and grounds an LLM answer in the |
| 168 | +reranked passages. |
| 169 | + |
| 170 | +## Where to next |
| 171 | + |
| 172 | +- [LLMs](llms.md) for the completion side that consumes retrieved context. |
| 173 | +- [Observability](observability.md) for what the embedding and rerank |
| 174 | + spans carry. |
| 175 | +- The [`openarmature.retrieval` reference](../reference/retrieval.md) for |
| 176 | + the full type surface. |
0 commit comments